mirror of https://github.com/microsoft/autogen.git
autogen subpackage (#968)
* math utils in autogen * cleanup * code utils * remove check function from code response * comment out test * GPT-4 * increase request timeout * name * logging and error handling * better doc * doc * codegen optimized * GPT series * text * no demo example * math * import openai * import openai * azure model name * azure model name * openai version * generate assertion if necessary * condition to generate assertions * init region key * rename * comments about budget * prompt --------- Co-authored-by: Susan Xueqing Liu <liususan091219@users.noreply.github.com>
This commit is contained in:
Chi Wang
GitHub
parent
7f9402b8fd
commit
82f0a4309d
20
README.md
20
README.md
|
|
@ -23,9 +23,9 @@
|
|||
## What is FLAML
|
||||
FLAML is a lightweight Python library that finds accurate machine
|
||||
learning models automatically, efficiently and economically. It frees users from selecting
|
||||
models and hyperparameters for each model. It can also be used to tune generic hyperparameters for large language models (LLM), MLOps/LMOps workflows, pipelines, mathematical/statistical models, algorithms, computing experiments, software configurations and so on.
|
||||
models and hyperparameters for each model. It can also be used to tune generic hyperparameters for foundation models, MLOps/LMOps workflows, pipelines, mathematical/statistical models, algorithms, computing experiments, software configurations and so on.
|
||||
|
||||
1. For common machine learning or AI tasks like classification, regression, and generation, it quickly finds quality models for user-provided data with low computational resources. It supports both classical machine learning models and deep neural networks, including large language models such as the OpenAI GPT-3 models.
|
||||
1. For common machine learning or AI tasks like classification, regression, and generation, it quickly finds quality models for user-provided data with low computational resources. It supports both classical machine learning models and deep neural networks, including foundation models such as the GPT series.
|
||||
1. It is easy to customize or extend. Users can find their desired customizability from a smooth range: minimal customization (computational resource budget), medium customization (e.g., scikit-style learner, search space and metric), or full customization (arbitrary training and evaluation code).
|
||||
1. It supports fast automatic tuning, capable of handling complex constraints/guidance/early stopping. FLAML is powered by a new, [cost-effective
|
||||
hyperparameter optimization](https://microsoft.github.io/FLAML/docs/Use-Cases/Tune-User-Defined-Function/#hyperparameter-optimization-algorithm)
|
||||
|
|
@ -95,6 +95,22 @@ estimator = LGBMRegressor()
|
|||
estimator.fit(X_train, y_train)
|
||||
```
|
||||
|
||||
* (New) You can optimize [generations](https://microsoft.github.io/FLAML/docs/Use-Cases/Auto-Generation) by ChatGPT or GPT-4 etc. with your own tuning data, success metrics and budgets.
|
||||
|
||||
```python
|
||||
from flaml import oai
|
||||
|
||||
config, analysis = oai.Completion.tune(
|
||||
data=tune_data,
|
||||
metric="success",
|
||||
mode="max",
|
||||
eval_func=eval_func,
|
||||
inference_budget=0.05,
|
||||
optimization_budget=3,
|
||||
num_samples=-1,
|
||||
)
|
||||
```
|
||||
|
||||
## Documentation
|
||||
|
||||
You can find a detailed documentation about FLAML [here](https://microsoft.github.io/FLAML/) where you can find the API documentation, use cases and examples.
|
||||
|
|
|
|||
|
|
@ -2,7 +2,7 @@ import logging
|
|||
from flaml.automl import AutoML, logger_formatter
|
||||
from flaml.tune.searcher import CFO, BlendSearch, FLOW2, BlendSearchTuner, RandomSearch
|
||||
from flaml.onlineml.autovw import AutoVW
|
||||
from flaml.integrations import oai
|
||||
from flaml.autogen import oai
|
||||
from flaml.version import __version__
|
||||
|
||||
|
||||
|
|
|
|||
|
|
@ -0,0 +1,181 @@
|
|||
import signal
|
||||
import subprocess
|
||||
import sys
|
||||
from typing import List, Dict, Tuple, Optional, Union, Callable
|
||||
from flaml import oai
|
||||
|
||||
|
||||
def timeout_handler(signum, frame):
|
||||
raise TimeoutError("Timed out!")
|
||||
|
||||
|
||||
def execute_code(code: str, max_exec_time: Optional[int] = 3):
|
||||
signal.signal(signal.SIGALRM, timeout_handler)
|
||||
code = code.strip()
|
||||
with open("codetest.py", "w") as fout:
|
||||
fout.write(code)
|
||||
try:
|
||||
signal.alarm(max_exec_time)
|
||||
result = subprocess.run(
|
||||
[sys.executable, "codetest.py"],
|
||||
stdout=subprocess.DEVNULL,
|
||||
stderr=subprocess.PIPE,
|
||||
)
|
||||
signal.alarm(0)
|
||||
except TimeoutError:
|
||||
return 0
|
||||
return int(result.returncode == 0)
|
||||
|
||||
|
||||
def generate_assertions(
|
||||
definition: str, model: Optional[str] = "gpt-3.5-turbo"
|
||||
) -> Tuple[str, float]:
|
||||
"""Generate assertions for a function.
|
||||
|
||||
Args:
|
||||
definition (str): The function definition, including the signature and docstr.
|
||||
model (str): The model used for generation.
|
||||
|
||||
Returns:
|
||||
str: The generated assertions.
|
||||
float: The cost of the generation.
|
||||
"""
|
||||
prompt = """Given the signature and docstring, write the exactly same number of assertion(s) for the provided example(s) in the docstring, without assertion messages.
|
||||
|
||||
func signature:
|
||||
{definition}
|
||||
assertions:"""
|
||||
response = oai.Completion.create(
|
||||
{"definition": definition},
|
||||
model=model,
|
||||
prompt=prompt,
|
||||
max_tokens=256,
|
||||
stop="\n\n",
|
||||
)
|
||||
cost = oai.Completion.cost(model, response)
|
||||
assertions = oai.Completion.extract_text(response)[0]
|
||||
return assertions, cost
|
||||
|
||||
|
||||
def _remove_check(response):
|
||||
"""Remove the check function from the response."""
|
||||
# find the position of the check function
|
||||
pos = response.find("def check(")
|
||||
if pos == -1:
|
||||
return response
|
||||
return response[:pos]
|
||||
|
||||
|
||||
def eval_function_completions(
|
||||
responses: List[str],
|
||||
definition: str,
|
||||
test: Optional[str] = None,
|
||||
entry_point: Optional[str] = None,
|
||||
assertions: Optional[Union[str, Callable[[str], Tuple[str, float]]]] = None,
|
||||
) -> Dict:
|
||||
"""Select a response from a list of responses for the function completion task (using generated assertions), and/or evaluate if the task is successful using a gold test.
|
||||
|
||||
Args:
|
||||
responses (list): The list of responses.
|
||||
definition (str): The input definition.
|
||||
test (Optional, str): The test code.
|
||||
entry_point (Optional, str): The name of the function.
|
||||
assertions (Optional, str or Callable): The assertion code which serves as a filter of the responses, or an assertion generator.
|
||||
When provided, only the responses that pass the assertions will be considered for the actual test (if provided).
|
||||
|
||||
Returns:
|
||||
dict: The success metrics.
|
||||
"""
|
||||
n = len(responses)
|
||||
if assertions is None:
|
||||
# no assertion filter
|
||||
success_list = []
|
||||
for i in range(n):
|
||||
response = _remove_check(responses[i])
|
||||
code = (
|
||||
f"{response}\n{test}\ncheck({entry_point})"
|
||||
if response.startswith("def")
|
||||
else f"{definition}{response}\n{test}\ncheck({entry_point})"
|
||||
)
|
||||
success = execute_code(code)
|
||||
success_list.append(success)
|
||||
return {
|
||||
"expected_success": 1 - pow(1 - sum(success_list) / n, n),
|
||||
"success": any(s for s in success_list),
|
||||
}
|
||||
if callable(assertions) and n > 1:
|
||||
# assertion generator
|
||||
assertions, gen_cost = assertions(definition)
|
||||
else:
|
||||
gen_cost = 0
|
||||
if n > 1 or test is None:
|
||||
for i in range(n):
|
||||
response = responses[i] = _remove_check(responses[i])
|
||||
code = (
|
||||
f"{response}\n{assertions}"
|
||||
if response.startswith("def")
|
||||
else f"{definition}{response}\n{assertions}"
|
||||
)
|
||||
succeed_assertions = execute_code(code)
|
||||
if succeed_assertions:
|
||||
break
|
||||
else:
|
||||
# just test, no need to check assertions
|
||||
succeed_assertions = False
|
||||
i, response = 0, responses[0]
|
||||
if test is None:
|
||||
# no test code
|
||||
return {
|
||||
"index_selected": i,
|
||||
"succeed_assertions": succeed_assertions,
|
||||
"gen_cost": gen_cost,
|
||||
"assertions": assertions,
|
||||
}
|
||||
code_test = (
|
||||
f"{response}\n{test}\ncheck({entry_point})"
|
||||
if response.startswith("def")
|
||||
else f"{definition}{response}\n{test}\ncheck({entry_point})"
|
||||
)
|
||||
success = execute_code(code_test)
|
||||
return {
|
||||
"index_selected": i,
|
||||
"succeed_assertions": succeed_assertions,
|
||||
"success": success,
|
||||
"gen_cost": gen_cost,
|
||||
"assertions": assertions,
|
||||
}
|
||||
|
||||
|
||||
def implement(
|
||||
definition: str,
|
||||
configs: List[Dict],
|
||||
assertions: Optional[
|
||||
Union[str, Callable[[str], Tuple[str, float]]]
|
||||
] = generate_assertions,
|
||||
) -> Tuple[str, float]:
|
||||
"""Implement a function from a definition.
|
||||
|
||||
Args:
|
||||
definition (str): The function definition, including the signature and docstr.
|
||||
configs (list): The list of configurations for completion.
|
||||
assertions (Optional, str or Callable): The assertion code which serves as a filter of the responses, or an assertion generator.
|
||||
|
||||
Returns:
|
||||
str: The implementation.
|
||||
float: The cost of the implementation.
|
||||
int: The index of the configuration which generates the implementation.
|
||||
"""
|
||||
cost = 0
|
||||
if len(configs) > 1 and callable(assertions):
|
||||
assertions, cost = assertions(definition)
|
||||
for i, config in enumerate(configs):
|
||||
response = oai.Completion.create({"definition": definition}, **config)
|
||||
cost += oai.Completion.cost(config["model"], response)
|
||||
responses = oai.Completion.extract_text(response)
|
||||
metrics = eval_function_completions(
|
||||
responses, definition, assertions=assertions
|
||||
)
|
||||
assertions = metrics["assertions"]
|
||||
cost += metrics["gen_cost"]
|
||||
if metrics["succeed_assertions"] or i == len(configs) - 1:
|
||||
return responses[metrics["index_selected"]], cost, i
|
||||
|
|
@ -0,0 +1,312 @@
|
|||
from typing import Optional
|
||||
|
||||
|
||||
def remove_boxed(string: str) -> Optional[str]:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Extract the text within a \\boxed{...} environment.
|
||||
Example:
|
||||
>>> remove_boxed(\\boxed{\\frac{2}{3}})
|
||||
\\frac{2}{3}
|
||||
"""
|
||||
left = "\\boxed{"
|
||||
try:
|
||||
assert string[: len(left)] == left
|
||||
assert string[-1] == "}"
|
||||
return string[len(left) : -1]
|
||||
except Exception:
|
||||
return None
|
||||
|
||||
|
||||
def last_boxed_only_string(string: str) -> Optional[str]:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Extract the last \\boxed{...} or \\fbox{...} element from a string.
|
||||
"""
|
||||
idx = string.rfind("\\boxed")
|
||||
if idx < 0:
|
||||
idx = string.rfind("\\fbox")
|
||||
if idx < 0:
|
||||
return None
|
||||
|
||||
i = idx
|
||||
right_brace_idx = None
|
||||
num_left_braces_open = 0
|
||||
while i < len(string):
|
||||
if string[i] == "{":
|
||||
num_left_braces_open += 1
|
||||
if string[i] == "}":
|
||||
num_left_braces_open -= 1
|
||||
if num_left_braces_open == 0:
|
||||
right_brace_idx = i
|
||||
break
|
||||
i += 1
|
||||
|
||||
if right_brace_idx is None:
|
||||
retval = None
|
||||
else:
|
||||
retval = string[idx : right_brace_idx + 1]
|
||||
|
||||
return retval
|
||||
|
||||
|
||||
def _fix_fracs(string: str) -> str:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Reformat fractions.
|
||||
Examples:
|
||||
>>> _fix_fracs("\\frac1b")
|
||||
\frac{1}{b}
|
||||
>>> _fix_fracs("\\frac12")
|
||||
\frac{1}{2}
|
||||
>>> _fix_fracs("\\frac1{72}")
|
||||
\frac{1}{72}
|
||||
"""
|
||||
substrs = string.split("\\frac")
|
||||
new_str = substrs[0]
|
||||
if len(substrs) > 1:
|
||||
substrs = substrs[1:]
|
||||
for substr in substrs:
|
||||
new_str += "\\frac"
|
||||
if substr[0] == "{":
|
||||
new_str += substr
|
||||
else:
|
||||
try:
|
||||
assert len(substr) >= 2
|
||||
except Exception:
|
||||
return string
|
||||
a = substr[0]
|
||||
b = substr[1]
|
||||
if b != "{":
|
||||
if len(substr) > 2:
|
||||
post_substr = substr[2:]
|
||||
new_str += "{" + a + "}{" + b + "}" + post_substr
|
||||
else:
|
||||
new_str += "{" + a + "}{" + b + "}"
|
||||
else:
|
||||
if len(substr) > 2:
|
||||
post_substr = substr[2:]
|
||||
new_str += "{" + a + "}" + b + post_substr
|
||||
else:
|
||||
new_str += "{" + a + "}" + b
|
||||
string = new_str
|
||||
return string
|
||||
|
||||
|
||||
def _fix_a_slash_b(string: str) -> str:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Reformat fractions formatted as a/b to \\frac{a}{b}.
|
||||
Example:
|
||||
>>> _fix_a_slash_b("2/3")
|
||||
\frac{2}{3}
|
||||
"""
|
||||
if len(string.split("/")) != 2:
|
||||
return string
|
||||
a_str = string.split("/")[0]
|
||||
b_str = string.split("/")[1]
|
||||
try:
|
||||
a = int(a_str)
|
||||
b = int(b_str)
|
||||
assert string == "{}/{}".format(a, b)
|
||||
new_string = "\\frac{" + str(a) + "}{" + str(b) + "}"
|
||||
return new_string
|
||||
except Exception:
|
||||
return string
|
||||
|
||||
|
||||
def _remove_right_units(string: str) -> str:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Remove units (on the right).
|
||||
"\\text{ " only ever occurs (at least in the val set) when describing units.
|
||||
"""
|
||||
if "\\text{ " in string:
|
||||
splits = string.split("\\text{ ")
|
||||
assert len(splits) == 2
|
||||
return splits[0]
|
||||
else:
|
||||
return string
|
||||
|
||||
|
||||
def _fix_sqrt(string: str) -> str:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Reformat square roots.
|
||||
Example:
|
||||
>>> _fix_sqrt("\\sqrt3")
|
||||
\\sqrt{3}
|
||||
"""
|
||||
if "\\sqrt" not in string:
|
||||
return string
|
||||
splits = string.split("\\sqrt")
|
||||
new_string = splits[0]
|
||||
for split in splits[1:]:
|
||||
if split[0] != "{":
|
||||
a = split[0]
|
||||
new_substr = "\\sqrt{" + a + "}" + split[1:]
|
||||
else:
|
||||
new_substr = "\\sqrt" + split
|
||||
new_string += new_substr
|
||||
return new_string
|
||||
|
||||
|
||||
def _strip_string(string: str) -> str:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Apply the reformatting helper functions above.
|
||||
"""
|
||||
# linebreaks
|
||||
string = string.replace("\n", "")
|
||||
# print(string)
|
||||
|
||||
# remove inverse spaces
|
||||
string = string.replace("\\!", "")
|
||||
# print(string)
|
||||
|
||||
# replace \\ with \
|
||||
string = string.replace("\\\\", "\\")
|
||||
# print(string)
|
||||
|
||||
# replace tfrac and dfrac with frac
|
||||
string = string.replace("tfrac", "frac")
|
||||
string = string.replace("dfrac", "frac")
|
||||
# print(string)
|
||||
|
||||
# remove \left and \right
|
||||
string = string.replace("\\left", "")
|
||||
string = string.replace("\\right", "")
|
||||
# print(string)
|
||||
|
||||
# Remove circ (degrees)
|
||||
string = string.replace("^{\\circ}", "")
|
||||
string = string.replace("^\\circ", "")
|
||||
|
||||
# remove dollar signs
|
||||
string = string.replace("\\$", "")
|
||||
|
||||
# remove units (on the right)
|
||||
string = _remove_right_units(string)
|
||||
|
||||
# remove percentage
|
||||
string = string.replace("\\%", "")
|
||||
string = string.replace("%", "")
|
||||
|
||||
# " 0." equivalent to " ." and "{0." equivalent to "{." Alternatively, add "0" if "." is the start of the string
|
||||
string = string.replace(" .", " 0.")
|
||||
string = string.replace("{.", "{0.")
|
||||
# if empty, return empty string
|
||||
if len(string) == 0:
|
||||
return string
|
||||
if string[0] == ".":
|
||||
string = "0" + string
|
||||
|
||||
# to consider: get rid of e.g. "k = " or "q = " at beginning
|
||||
if len(string.split("=")) == 2:
|
||||
if len(string.split("=")[0]) <= 2:
|
||||
string = string.split("=")[1]
|
||||
|
||||
# fix sqrt3 --> sqrt{3}
|
||||
string = _fix_sqrt(string)
|
||||
|
||||
# remove spaces
|
||||
string = string.replace(" ", "")
|
||||
|
||||
# \frac1b or \frac12 --> \frac{1}{b} and \frac{1}{2}, etc.
|
||||
# Even works with \frac1{72} (but not \frac{72}1).
|
||||
# Also does a/b --> \\frac{a}{b}
|
||||
string = _fix_fracs(string)
|
||||
|
||||
# manually change 0.5 --> \frac{1}{2}
|
||||
if string == "0.5":
|
||||
string = "\\frac{1}{2}"
|
||||
|
||||
# NOTE: X/Y changed to \frac{X}{Y} in dataset, but in simple cases fix in case the model output is X/Y
|
||||
string = _fix_a_slash_b(string)
|
||||
|
||||
return string
|
||||
|
||||
|
||||
def get_answer(solution: Optional[str]) -> Optional[str]:
|
||||
if solution is None:
|
||||
return None
|
||||
last_boxed = last_boxed_only_string(solution)
|
||||
if last_boxed is None:
|
||||
return None
|
||||
answer = remove_boxed(last_boxed)
|
||||
if answer is None:
|
||||
return None
|
||||
return answer
|
||||
|
||||
|
||||
def is_equiv(str1: Optional[str], str2: Optional[str]) -> float:
|
||||
"""Returns (as a float) whether two strings containing math are equivalent up to differences of formatting in
|
||||
- units
|
||||
- fractions
|
||||
- square roots
|
||||
- superfluous LaTeX.
|
||||
Source: https://github.com/hendrycks/math
|
||||
"""
|
||||
if str1 is None and str2 is None:
|
||||
print("WARNING: Both None")
|
||||
return 1.0
|
||||
if str1 is None or str2 is None:
|
||||
return 0.0
|
||||
|
||||
try:
|
||||
ss1 = _strip_string(str1)
|
||||
ss2 = _strip_string(str2)
|
||||
return float(ss1 == ss2)
|
||||
except Exception:
|
||||
return float(str1 == str2)
|
||||
|
||||
|
||||
def is_equiv_chain_of_thought(str1: str, str2: str) -> float:
|
||||
"""Strips the solution first before calling `is_equiv`."""
|
||||
ans1 = get_answer(str1)
|
||||
ans2 = get_answer(str2)
|
||||
|
||||
return is_equiv(ans1, ans2)
|
||||
|
||||
|
||||
def voting_counts(responses):
|
||||
answers = {}
|
||||
for i in range(len(responses)):
|
||||
equiv = i
|
||||
if get_answer(responses[i]) is None:
|
||||
# ignore None answers
|
||||
continue
|
||||
for j in answers:
|
||||
if is_equiv_chain_of_thought(responses[i], responses[j]):
|
||||
equiv = j
|
||||
break
|
||||
if equiv in answers:
|
||||
answers[equiv] += 1
|
||||
else:
|
||||
answers[equiv] = 1
|
||||
return answers
|
||||
|
||||
|
||||
def eval_math_responses(responses, solution=None, **args):
|
||||
"""Select a response for a math problem using voting, and check if the response is correct if the solution is provided.
|
||||
|
||||
Args:
|
||||
responses (list): The list of responses.
|
||||
solution (str): The canonical solution.
|
||||
|
||||
Returns:
|
||||
dict: The success metrics.
|
||||
"""
|
||||
success_list = []
|
||||
n = len(responses)
|
||||
if solution is not None:
|
||||
for i in range(n):
|
||||
response = responses[i]
|
||||
succeed = is_equiv_chain_of_thought(response, solution)
|
||||
success_list.append(succeed)
|
||||
# voting
|
||||
answers = voting_counts(responses)
|
||||
# find the answer with highest votes in answers
|
||||
answer, votes = max(answers.items(), key=lambda x: x[1], default=(0, 0))
|
||||
# check if the answer is correct
|
||||
success_vote = is_equiv_chain_of_thought(responses[answer], solution)
|
||||
return {
|
||||
"expected_success": 1 - pow(1 - sum(success_list) / n, n),
|
||||
"success": any(s for s in success_list),
|
||||
"success_vote": success_vote,
|
||||
"voted_answer": responses[answer],
|
||||
"votes": votes,
|
||||
}
|
||||
|
|
@ -0,0 +1,3 @@
|
|||
from flaml.autogen.oai.completion import Completion, ChatCompletion
|
||||
|
||||
__all__ = ["Completion", "ChatCompletion"]
|
||||
|
|
@ -2,7 +2,10 @@ from time import sleep
|
|||
import logging
|
||||
import numpy as np
|
||||
import time
|
||||
from typing import List
|
||||
import sys
|
||||
from flaml import tune, BlendSearch
|
||||
from flaml.automl.logger import logger_formatter
|
||||
|
||||
try:
|
||||
import openai
|
||||
|
|
@ -22,6 +25,11 @@ except ImportError:
|
|||
"please install flaml[openai] option to use the flaml.oai subpackage."
|
||||
)
|
||||
logger = logging.getLogger(__name__)
|
||||
if not logger.handlers:
|
||||
# Add the console handler.
|
||||
_ch = logging.StreamHandler(stream=sys.stdout)
|
||||
_ch.setFormatter(logger_formatter)
|
||||
logger.addHandler(_ch)
|
||||
|
||||
|
||||
def get_key(config):
|
||||
|
|
@ -50,6 +58,7 @@ class Completion:
|
|||
chat_models = {
|
||||
"gpt-3.5-turbo",
|
||||
"gpt-3.5-turbo-0301",
|
||||
"gpt-35-turbo",
|
||||
"gpt-4",
|
||||
"gpt-4-32k",
|
||||
"gpt-4-32k-0314",
|
||||
|
|
@ -67,6 +76,7 @@ class Completion:
|
|||
"text-davinci-003": 0.02,
|
||||
"gpt-3.5-turbo": 0.002,
|
||||
"gpt-3.5-turbo-0301": 0.002,
|
||||
"gpt-35-turbo": 0.002,
|
||||
"gpt-4": (0.03, 0.06),
|
||||
"gpt-4-0314": (0.03, 0.06),
|
||||
"gpt-4-32k": (0.06, 0.12),
|
||||
|
|
@ -95,12 +105,13 @@ class Completion:
|
|||
}
|
||||
|
||||
seed = 41
|
||||
cache_path = f".cache/{seed}"
|
||||
# retry after this many seconds
|
||||
retry_time = 10
|
||||
# fail a request after hitting RateLimitError for this many seconds
|
||||
retry_timeout = 60
|
||||
retry_timeout = 120
|
||||
# time out for request to openai server
|
||||
request_timeout = 30
|
||||
request_timeout = 60
|
||||
|
||||
openai_completion_class = not ERROR and openai.Completion
|
||||
_total_cost = 0
|
||||
|
|
@ -156,14 +167,18 @@ class Completion:
|
|||
# retry after retry_time seconds
|
||||
if time.time() - start_time + cls.retry_time < cls.retry_timeout:
|
||||
logger.info(f"retrying in {cls.retry_time} seconds...", exc_info=1)
|
||||
elif not eval_only:
|
||||
elif eval_only:
|
||||
raise
|
||||
else:
|
||||
break
|
||||
sleep(cls.retry_time)
|
||||
except InvalidRequestError:
|
||||
if "azure" == openai.api_type and "model" in config:
|
||||
# azure api uses "engine" instead of "model"
|
||||
config = config.copy()
|
||||
config["engine"] = config.pop("model")
|
||||
config["engine"] = config.pop("model").replace(
|
||||
"gpt-3.5-turbo", "gpt-35-turbo"
|
||||
)
|
||||
else:
|
||||
raise
|
||||
logger.warning(
|
||||
|
|
@ -219,6 +234,13 @@ class Completion:
|
|||
num_completions, invalid_n.get(max_tokens, np.inf)
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def _pop_subspace(cls, config):
|
||||
if "subspace" in config:
|
||||
config = config.copy()
|
||||
config.update(config.pop("subspace"))
|
||||
return config
|
||||
|
||||
@classmethod
|
||||
def _get_prompt_messages_from_config(cls, model, config):
|
||||
prompt, messages = None, None
|
||||
|
|
@ -254,6 +276,7 @@ class Completion:
|
|||
"""
|
||||
cost = 0
|
||||
data = cls.data
|
||||
config = cls._pop_subspace(config)
|
||||
model = config["model"]
|
||||
data_length = len(data)
|
||||
price = cls.price1K.get(model)
|
||||
|
|
@ -300,8 +323,10 @@ class Completion:
|
|||
start_n = max_valid_n + 1
|
||||
else:
|
||||
start_n = config_n
|
||||
region_key = None
|
||||
params = config.copy()
|
||||
params["stop"] = stop
|
||||
if "stop" in config:
|
||||
params["stop"] = stop
|
||||
temperature_or_top_p = params.pop("temperature_or_top_p", None)
|
||||
if temperature_or_top_p:
|
||||
params.update(temperature_or_top_p)
|
||||
|
|
@ -329,11 +354,7 @@ class Completion:
|
|||
result["cost"] = cost
|
||||
return result
|
||||
# evaluate the quality of the responses
|
||||
responses = (
|
||||
[r["message"]["content"].rstrip() for r in response["choices"]]
|
||||
if model in cls.chat_models
|
||||
else [r["text"].rstrip() for r in response["choices"]]
|
||||
)
|
||||
responses = cls.extract_text(response)
|
||||
usage = response["usage"]
|
||||
n_input_tokens = usage["prompt_tokens"]
|
||||
n_output_tokens = usage.get("completion_tokens", 0)
|
||||
|
|
@ -491,11 +512,12 @@ class Completion:
|
|||
```
|
||||
|
||||
log_file_name (str, optional): The log file.
|
||||
inference_budget (float, optional): The inference budget.
|
||||
optimization_budget (float, optional): The optimization budget.
|
||||
inference_budget (float, optional): The inference budget, dollar per instance.
|
||||
optimization_budget (float, optional): The optimization budget, dollar in total.
|
||||
num_samples (int, optional): The number of samples to evaluate.
|
||||
-1 means no hard restriction in the number of trials
|
||||
and the actual number is decided by optimization_budget. Defaults to 1.
|
||||
logging_level (optional): logging level. Defaults to logging.WARNING.
|
||||
**config (dict): The search space to update over the default search.
|
||||
For prompt, please provide a string/Callable or a list of strings/Callables.
|
||||
- If prompt is provided for chat models, it will be converted to messages under role "user".
|
||||
|
|
@ -570,22 +592,38 @@ class Completion:
|
|||
cls.data = data
|
||||
cls.avg_input_tokens = None
|
||||
|
||||
search_alg = BlendSearch(
|
||||
cost_attr="cost",
|
||||
cost_budget=optimization_budget,
|
||||
metric=metric,
|
||||
mode=mode,
|
||||
space=space,
|
||||
)
|
||||
space_model = space["model"]
|
||||
if not isinstance(space_model, str) and len(space_model) > 1:
|
||||
# make a hierarchical search space
|
||||
subspace = {}
|
||||
if "max_tokens" in space:
|
||||
subspace["max_tokens"] = space.pop("max_tokens")
|
||||
if "temperature_or_top_p" in space:
|
||||
subspace["temperature_or_top_p"] = space.pop("temperature_or_top_p")
|
||||
if "best_of" in space:
|
||||
subspace["best_of"] = space.pop("best_of")
|
||||
if "n" in space:
|
||||
subspace["n"] = space.pop("n")
|
||||
choices = []
|
||||
for model in space["model"]:
|
||||
choices.append({"model": model, **subspace})
|
||||
space["subspace"] = tune.choice(choices)
|
||||
space.pop("model")
|
||||
# start all the models with the same hp config
|
||||
search_alg = BlendSearch(
|
||||
cost_attr="cost",
|
||||
cost_budget=optimization_budget,
|
||||
metric=metric,
|
||||
mode=mode,
|
||||
space=space,
|
||||
)
|
||||
config0 = search_alg.suggest("t0")
|
||||
points_to_evaluate = [config0]
|
||||
for model in space_model:
|
||||
if model != config0["model"]:
|
||||
if model != config0["subspace"]["model"]:
|
||||
point = config0.copy()
|
||||
point["model"] = model
|
||||
point["subspace"] = point["subspace"].copy()
|
||||
point["subspace"]["model"] = model
|
||||
points_to_evaluate.append(point)
|
||||
search_alg = BlendSearch(
|
||||
cost_attr="cost",
|
||||
|
|
@ -595,6 +633,15 @@ class Completion:
|
|||
space=space,
|
||||
points_to_evaluate=points_to_evaluate,
|
||||
)
|
||||
else:
|
||||
search_alg = BlendSearch(
|
||||
cost_attr="cost",
|
||||
cost_budget=optimization_budget,
|
||||
metric=metric,
|
||||
mode=mode,
|
||||
space=space,
|
||||
)
|
||||
old_level = logger.getEffectiveLevel()
|
||||
logger.setLevel(logging_level)
|
||||
with diskcache.Cache(cls.cache_path) as cls._cache:
|
||||
analysis = tune.run(
|
||||
|
|
@ -605,7 +652,7 @@ class Completion:
|
|||
verbose=3,
|
||||
)
|
||||
config = analysis.best_config
|
||||
params = config.copy()
|
||||
params = cls._pop_subspace(config)
|
||||
if cls._prompts:
|
||||
params["prompt"] = cls._prompts[config["prompt"]]
|
||||
else:
|
||||
|
|
@ -615,6 +662,7 @@ class Completion:
|
|||
temperature_or_top_p = params.pop("temperature_or_top_p", None)
|
||||
if temperature_or_top_p:
|
||||
params.update(temperature_or_top_p)
|
||||
logger.setLevel(old_level)
|
||||
return params, analysis
|
||||
|
||||
@classmethod
|
||||
|
|
@ -636,12 +684,14 @@ class Completion:
|
|||
if ERROR:
|
||||
raise ERROR
|
||||
params = cls._construct_params(context, config)
|
||||
if use_cache:
|
||||
with diskcache.Cache(cls.cache_path) as cls._cache:
|
||||
return cls._get_response(params)
|
||||
return cls.openai_completion_class.create(
|
||||
request_timeout=cls.request_timeout, **params
|
||||
)
|
||||
if not use_cache:
|
||||
return cls._get_response(params, eval_only=True, use_cache=False)
|
||||
seed = cls.seed
|
||||
if "seed" in params:
|
||||
cls.set_cache(params.pop("seed"))
|
||||
with diskcache.Cache(cls.cache_path) as cls._cache:
|
||||
cls.set_cache(seed)
|
||||
return cls._get_response(params, eval_only=True)
|
||||
|
||||
@classmethod
|
||||
def _construct_params(cls, data_instance, config, prompt=None, messages=None):
|
||||
|
|
@ -698,8 +748,7 @@ class Completion:
|
|||
use_cache=True,
|
||||
agg_method="avg",
|
||||
return_responses_and_per_instance_result=False,
|
||||
seed=41,
|
||||
cache_path=".cache",
|
||||
logging_level=logging.WARNING,
|
||||
):
|
||||
"""Evaluate the responses created with the config for the OpenAI API call.
|
||||
|
||||
|
|
@ -750,54 +799,45 @@ class Completion:
|
|||
|
||||
return_responses_and_per_instance_result (bool): Whether to also return responses
|
||||
and per instance results in addition to the aggregated results.
|
||||
seed (int): Random seed for the evaluation. Defaults to 41.
|
||||
cache_path (str): Path to the cache directory. Defaults to '.cache'.
|
||||
If a cache directory does not exist, it will be created, otherwise use the existing one.
|
||||
logging_level (optional): logging level. Defaults to logging.WARNING.
|
||||
|
||||
Returns:
|
||||
None in case of rate limit error or when a valid eval_func is not provided in either test or tune;
|
||||
None when no valid eval_func is provided in either test or tune;
|
||||
Otherwise, a dict of aggregated results, responses and per instance results if `return_responses_and_per_instance_result` is True;
|
||||
Otherwise, a dict of aggregated results (responses and per instance results are not returned).
|
||||
"""
|
||||
model = config["model"]
|
||||
result_agg, responses_list, result_list = {}, [], []
|
||||
metric_keys = None
|
||||
cls.set_cache(seed, cache_path)
|
||||
with diskcache.Cache(cls.cache_path) as cls._cache:
|
||||
for i, data_i in enumerate(data):
|
||||
logger.info(f"evaluating data instance {i}")
|
||||
params = cls._construct_params(data_i, config)
|
||||
response = cls._get_response(
|
||||
params, eval_only=True, use_cache=use_cache
|
||||
cost = 0
|
||||
model = config["model"]
|
||||
old_level = logger.getEffectiveLevel()
|
||||
logger.setLevel(logging_level)
|
||||
for i, data_i in enumerate(data):
|
||||
logger.info(f"evaluating data instance {i}")
|
||||
response = cls.create(data_i, use_cache, **config)
|
||||
cost += cls.cost(model, response)
|
||||
# evaluate the quality of the responses
|
||||
responses = cls.extract_text(response)
|
||||
if eval_func is not None:
|
||||
metrics = eval_func(responses, **data_i)
|
||||
elif hasattr(cls, "_eval_func"):
|
||||
metrics = cls._eval_func(responses, **data_i)
|
||||
else:
|
||||
logger.warning(
|
||||
"Please either provide a valid eval_func or do the test after the tune function is called."
|
||||
)
|
||||
if response == -1: # rate limit error, treat as invalid
|
||||
return None
|
||||
# evaluate the quality of the responses
|
||||
responses = (
|
||||
[r["message"]["content"].rstrip() for r in response["choices"]]
|
||||
if model in cls.chat_models
|
||||
else [r["text"].rstrip() for r in response["choices"]]
|
||||
)
|
||||
|
||||
if eval_func is not None:
|
||||
metrics = eval_func(responses, **data_i)
|
||||
elif hasattr(cls, "_eval_func"):
|
||||
metrics = cls._eval_func(responses, **data_i)
|
||||
else:
|
||||
logger.warning(
|
||||
"Please either provide a valid eval_func or do the test after the tune function is called"
|
||||
)
|
||||
return
|
||||
if not metric_keys:
|
||||
metric_keys = []
|
||||
for k in metrics.keys():
|
||||
try:
|
||||
_ = float(metrics[k])
|
||||
metric_keys.append(k)
|
||||
except ValueError:
|
||||
pass
|
||||
result_list.append(metrics)
|
||||
if return_responses_and_per_instance_result:
|
||||
responses_list.append(responses)
|
||||
return
|
||||
if not metric_keys:
|
||||
metric_keys = []
|
||||
for k in metrics.keys():
|
||||
try:
|
||||
_ = float(metrics[k])
|
||||
metric_keys.append(k)
|
||||
except ValueError:
|
||||
pass
|
||||
result_list.append(metrics)
|
||||
if return_responses_and_per_instance_result:
|
||||
responses_list.append(responses)
|
||||
if isinstance(agg_method, str):
|
||||
if agg_method in ["avg", "average"]:
|
||||
for key in metric_keys:
|
||||
|
|
@ -824,25 +864,57 @@ class Completion:
|
|||
"agg_method needs to be a string ('avg' or 'median'),\
|
||||
or a callable, or a dictionary of callable."
|
||||
)
|
||||
logger.setLevel(old_level)
|
||||
# should we also return the result_list and responses_list or not?
|
||||
if "cost" not in result_agg:
|
||||
result_agg["cost"] = cost
|
||||
if "inference_cost" not in result_agg:
|
||||
result_agg["inference_cost"] = cost / len(data)
|
||||
if return_responses_and_per_instance_result:
|
||||
return result_agg, result_list, responses_list
|
||||
else:
|
||||
return result_agg
|
||||
|
||||
@classmethod
|
||||
def cost(cls, model: str, response: dict):
|
||||
"""Compute the cost of a completion.
|
||||
|
||||
Args:
|
||||
model (str): The model name.
|
||||
response (dict): The response from OpenAI API.
|
||||
|
||||
Returns:
|
||||
The cost in USD.
|
||||
"""
|
||||
if model not in cls.price1K:
|
||||
raise ValueError(f"Unknown model: {model}")
|
||||
usage = response["usage"]
|
||||
n_input_tokens = usage["prompt_tokens"]
|
||||
n_output_tokens = usage.get("completion_tokens", 0)
|
||||
price1K = cls.price1K[model]
|
||||
if isinstance(price1K, tuple):
|
||||
return (price1K[0] * n_input_tokens + price1K[1] * n_output_tokens) / 1000
|
||||
return price1K * (n_input_tokens + n_output_tokens) / 1000
|
||||
|
||||
@classmethod
|
||||
def extract_text(cls, response: dict) -> List[str]:
|
||||
"""Extract the text from a completion response.
|
||||
|
||||
Args:
|
||||
response (dict): The response from OpenAI API.
|
||||
|
||||
Returns:
|
||||
A list of text in the responses.
|
||||
"""
|
||||
choices = response["choices"]
|
||||
if "text" in choices[0]:
|
||||
return [choice["text"] for choice in choices]
|
||||
return [choice["message"]["content"] for choice in choices]
|
||||
|
||||
|
||||
class ChatCompletion(Completion):
|
||||
"""A class for OpenAI API ChatCompletion."""
|
||||
|
||||
price1K = {
|
||||
"gpt-3.5-turbo": 0.002,
|
||||
"gpt-3.5-turbo-0301": 0.002,
|
||||
"gpt-4": (0.03, 0.06),
|
||||
"gpt-4-0314": (0.03, 0.06),
|
||||
"gpt-4-32k": (0.06, 0.12),
|
||||
"gpt-4-32k-0314": (0.06, 0.12),
|
||||
}
|
||||
|
||||
default_search_space = Completion.default_search_space.copy()
|
||||
default_search_space["model"] = tune.choice(["gpt-3.5-turbo", "gpt-4"])
|
||||
openai_completion_class = not ERROR and openai.ChatCompletion
|
||||
|
|
@ -1,3 +0,0 @@
|
|||
from flaml.integrations.oai.completion import Completion, ChatCompletion
|
||||
|
||||
__all__ = ["Completion", "ChatCompletion"]
|
||||
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
|
|
@ -0,0 +1,787 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"slideshow": {
|
||||
"slide_type": "slide"
|
||||
}
|
||||
},
|
||||
"source": [
|
||||
"Copyright (c) Microsoft Corporation. All rights reserved. \n",
|
||||
"\n",
|
||||
"Licensed under the MIT License.\n",
|
||||
"\n",
|
||||
"# Use FLAML to Optimize Code Generation Performance\n",
|
||||
"\n",
|
||||
"In this notebook, we optimize OpenAI models for code generation. We use [the HumanEval benchmark](https://huggingface.co/datasets/openai_humaneval) released by OpenAI for synthesizing programs from docstrings. \n",
|
||||
"\n",
|
||||
"## Requirements\n",
|
||||
"\n",
|
||||
"FLAML requires `Python>=3.7`. To run this notebook example, please install flaml with the [openai] option:\n",
|
||||
"```bash\n",
|
||||
"pip install flaml[openai]==1.2.0\n",
|
||||
"```"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-24T23:25:36.910966Z",
|
||||
"iopub.status.busy": "2023-02-24T23:25:36.910473Z",
|
||||
"iopub.status.idle": "2023-02-24T23:25:36.914554Z",
|
||||
"shell.execute_reply": "2023-02-24T23:25:36.914030Z"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# %pip install flaml[openai]==1.2.0 datasets"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"Set your OpenAI key:"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-24T23:25:36.917301Z",
|
||||
"iopub.status.busy": "2023-02-24T23:25:36.917011Z",
|
||||
"iopub.status.idle": "2023-02-24T23:25:36.923156Z",
|
||||
"shell.execute_reply": "2023-02-24T23:25:36.922619Z"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import os\n",
|
||||
"\n",
|
||||
"if \"OPENAI_API_KEY\" not in os.environ:\n",
|
||||
" os.environ[\"OPENAI_API_KEY\"] = \"<your OpenAI API key here>\""
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"If you use Azure OpenAI, uncomment the following:"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 3,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-24T23:25:36.925804Z",
|
||||
"iopub.status.busy": "2023-02-24T23:25:36.925423Z",
|
||||
"iopub.status.idle": "2023-02-24T23:25:36.928191Z",
|
||||
"shell.execute_reply": "2023-02-24T23:25:36.927673Z"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# import openai\n",
|
||||
"# openai.api_type = \"azure\"\n",
|
||||
"# openai.api_base = \"https://<your_endpoint>.openai.azure.com/\"\n",
|
||||
"# openai.api_version = \"2023-03-15-preview\" # change if necessary"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## Load dataset\n",
|
||||
"\n",
|
||||
"First, we load the humaneval dataset. The dataset contains 164 examples. In each example, the \"prompt\" is the prompt string for eliciting the code generation (renamed into \"definition\"), \"test\" is the Python code for unit test for the example, and \"entry_point\" is the function name to be tested."
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 4,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-24T23:25:36.931255Z",
|
||||
"iopub.status.busy": "2023-02-24T23:25:36.930838Z",
|
||||
"iopub.status.idle": "2023-02-24T23:25:39.148799Z",
|
||||
"shell.execute_reply": "2023-02-24T23:25:39.148113Z"
|
||||
}
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Found cached dataset openai_humaneval (/home/vscode/.cache/huggingface/datasets/openai_humaneval/openai_humaneval/1.0.0/2955cebd73602e828fa8c0a424c594e5fab4ec863b316ca98f3d8fdb6a626e75)\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"application/vnd.jupyter.widget-view+json": {
|
||||
"model_id": "1fdc8853bf2a4aecaa2cd024ad99b5a2",
|
||||
"version_major": 2,
|
||||
"version_minor": 0
|
||||
},
|
||||
"text/plain": [
|
||||
" 0%| | 0/1 [00:00<?, ?it/s]"
|
||||
]
|
||||
},
|
||||
"metadata": {},
|
||||
"output_type": "display_data"
|
||||
},
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Loading cached shuffled indices for dataset at /home/vscode/.cache/huggingface/datasets/openai_humaneval/openai_humaneval/1.0.0/2955cebd73602e828fa8c0a424c594e5fab4ec863b316ca98f3d8fdb6a626e75/cache-1e8448101c1b32e8.arrow\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"import datasets\n",
|
||||
"\n",
|
||||
"seed = 41\n",
|
||||
"data = datasets.load_dataset(\"openai_humaneval\")[\"test\"].shuffle(seed=seed)\n",
|
||||
"data = data.select(range(len(data))).rename_column(\"prompt\", \"definition\").remove_columns([\"task_id\", \"canonical_solution\"])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-24T23:25:39.164187Z",
|
||||
"iopub.status.busy": "2023-02-24T23:25:39.163867Z",
|
||||
"iopub.status.idle": "2023-02-24T23:25:39.169009Z",
|
||||
"shell.execute_reply": "2023-02-24T23:25:39.168427Z"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from flaml.autogen.code_utils import eval_function_completions, implement\n",
|
||||
"from flaml import oai"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"slideshow": {
|
||||
"slide_type": "slide"
|
||||
}
|
||||
},
|
||||
"source": [
|
||||
"The `implement` function will first generate assertion statements for a problem. Then, it uses the assertions to select the generated responses."
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-24T23:25:39.179030Z",
|
||||
"iopub.status.busy": "2023-02-24T23:25:39.178624Z",
|
||||
"iopub.status.idle": "2023-02-24T23:25:40.584410Z",
|
||||
"shell.execute_reply": "2023-02-24T23:25:40.583802Z"
|
||||
},
|
||||
"slideshow": {
|
||||
"slide_type": "slide"
|
||||
}
|
||||
},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Example 0, config 1, success 1\n",
|
||||
"Example 1, config 0, success 2\n",
|
||||
"Example 2, config 0, success 3\n",
|
||||
"Example 3, config 2, success 4\n",
|
||||
"Example 4, config 2, success 5\n",
|
||||
"Example 5, config 4, success 6\n",
|
||||
"Example 6, config 4, success 6\n",
|
||||
"Example 7, config 2, success 7\n",
|
||||
"Example 8, config 2, success 8\n",
|
||||
"Example 9, config 0, success 9\n",
|
||||
"Example 10, config 1, success 10\n",
|
||||
"Example 11, config 0, success 10\n",
|
||||
"Example 12, config 2, success 11\n",
|
||||
"Example 13, config 2, success 12\n",
|
||||
"Example 14, config 0, success 13\n",
|
||||
"Example 15, config 2, success 14\n",
|
||||
"Example 16, config 0, success 15\n",
|
||||
"Example 17, config 1, success 15\n",
|
||||
"Example 18, config 1, success 16\n",
|
||||
"Example 19, config 3, success 17\n",
|
||||
"Example 20, config 2, success 18\n",
|
||||
"Example 21, config 2, success 19\n",
|
||||
"Example 22, config 2, success 19\n",
|
||||
"Example 23, config 2, success 20\n",
|
||||
"Example 24, config 0, success 21\n",
|
||||
"Example 25, config 0, success 22\n",
|
||||
"Example 26, config 4, success 23\n",
|
||||
"Example 27, config 2, success 24\n",
|
||||
"Example 28, config 4, success 24\n",
|
||||
"Example 29, config 2, success 25\n",
|
||||
"Example 30, config 2, success 26\n",
|
||||
"Example 31, config 0, success 27\n",
|
||||
"Example 32, config 0, success 28\n",
|
||||
"Example 33, config 0, success 29\n",
|
||||
"Example 34, config 2, success 30\n",
|
||||
"Example 35, config 1, success 30\n",
|
||||
"Example 36, config 0, success 31\n",
|
||||
"Example 37, config 0, success 32\n",
|
||||
"Example 38, config 0, success 33\n",
|
||||
"Example 39, config 2, success 34\n",
|
||||
"Example 40, config 0, success 35\n",
|
||||
"Example 41, config 0, success 36\n",
|
||||
"Example 42, config 3, success 37\n",
|
||||
"Example 43, config 0, success 38\n",
|
||||
"Example 44, config 2, success 39\n",
|
||||
"Example 45, config 2, success 40\n",
|
||||
"Example 46, config 2, success 40\n",
|
||||
"Example 47, config 0, success 41\n",
|
||||
"Example 48, config 3, success 42\n",
|
||||
"Example 49, config 2, success 43\n",
|
||||
"Example 50, config 1, success 44\n",
|
||||
"Example 51, config 2, success 45\n",
|
||||
"Example 52, config 3, success 46\n",
|
||||
"Example 53, config 2, success 47\n",
|
||||
"Example 54, config 0, success 48\n",
|
||||
"Example 55, config 2, success 49\n",
|
||||
"Example 56, config 2, success 50\n",
|
||||
"Example 57, config 2, success 51\n",
|
||||
"Example 58, config 0, success 52\n",
|
||||
"Example 59, config 1, success 53\n",
|
||||
"Example 60, config 0, success 53\n",
|
||||
"Example 61, config 0, success 54\n",
|
||||
"Example 62, config 1, success 55\n",
|
||||
"Example 63, config 1, success 56\n",
|
||||
"Example 64, config 0, success 57\n",
|
||||
"Example 65, config 2, success 58\n",
|
||||
"Example 66, config 2, success 59\n",
|
||||
"Example 67, config 2, success 60\n",
|
||||
"Example 68, config 2, success 61\n",
|
||||
"Example 69, config 4, success 61\n",
|
||||
"Example 70, config 2, success 62\n",
|
||||
"Example 71, config 0, success 63\n",
|
||||
"Example 72, config 0, success 64\n",
|
||||
"Example 73, config 0, success 65\n",
|
||||
"Example 74, config 0, success 66\n",
|
||||
"Example 75, config 0, success 67\n",
|
||||
"Example 76, config 1, success 68\n",
|
||||
"Example 77, config 2, success 69\n",
|
||||
"Example 78, config 1, success 70\n",
|
||||
"Example 79, config 4, success 70\n",
|
||||
"Example 80, config 2, success 71\n",
|
||||
"Example 81, config 2, success 72\n",
|
||||
"Example 82, config 0, success 72\n",
|
||||
"Example 83, config 0, success 73\n",
|
||||
"Example 84, config 4, success 73\n",
|
||||
"Example 85, config 3, success 74\n",
|
||||
"Example 86, config 0, success 75\n",
|
||||
"Example 87, config 2, success 76\n",
|
||||
"Example 88, config 2, success 77\n",
|
||||
"Example 89, config 1, success 78\n",
|
||||
"Example 90, config 0, success 79\n",
|
||||
"Example 91, config 2, success 80\n",
|
||||
"Example 92, config 1, success 81\n",
|
||||
"Example 93, config 0, success 82\n",
|
||||
"Example 94, config 0, success 83\n",
|
||||
"Example 95, config 0, success 84\n",
|
||||
"Example 96, config 2, success 85\n",
|
||||
"Example 97, config 2, success 86\n",
|
||||
"Example 98, config 2, success 87\n",
|
||||
"Example 99, config 4, success 88\n",
|
||||
"Example 100, config 0, success 89\n",
|
||||
"Example 101, config 0, success 90\n",
|
||||
"Example 102, config 2, success 91\n",
|
||||
"Example 103, config 4, success 91\n",
|
||||
"Example 104, config 2, success 92\n",
|
||||
"Example 105, config 2, success 93\n",
|
||||
"Example 106, config 4, success 93\n",
|
||||
"Example 107, config 2, success 94\n",
|
||||
"Example 108, config 0, success 95\n",
|
||||
"Example 109, config 2, success 96\n",
|
||||
"Example 110, config 0, success 97\n",
|
||||
"Example 111, config 0, success 98\n",
|
||||
"Example 112, config 2, success 99\n",
|
||||
"Example 113, config 0, success 99\n",
|
||||
"Example 114, config 2, success 100\n",
|
||||
"Example 115, config 2, success 100\n",
|
||||
"Example 116, config 0, success 101\n",
|
||||
"Example 117, config 0, success 102\n",
|
||||
"Example 118, config 0, success 103\n",
|
||||
"Example 119, config 4, success 104\n",
|
||||
"Example 120, config 2, success 105\n",
|
||||
"Example 121, config 2, success 106\n",
|
||||
"Example 122, config 0, success 107\n",
|
||||
"Example 123, config 2, success 108\n",
|
||||
"Example 124, config 1, success 109\n",
|
||||
"Example 125, config 0, success 110\n",
|
||||
"Example 126, config 1, success 111\n",
|
||||
"Example 127, config 4, success 111\n",
|
||||
"Example 128, config 2, success 112\n",
|
||||
"Example 129, config 2, success 113\n",
|
||||
"Example 130, config 0, success 114\n",
|
||||
"Example 131, config 2, success 115\n",
|
||||
"Example 132, config 0, success 116\n",
|
||||
"Example 133, config 2, success 117\n",
|
||||
"Example 134, config 1, success 118\n",
|
||||
"Example 135, config 1, success 119\n",
|
||||
"Example 136, config 0, success 120\n",
|
||||
"Example 137, config 0, success 121\n",
|
||||
"Example 138, config 2, success 122\n",
|
||||
"Example 139, config 2, success 123\n",
|
||||
"Example 140, config 2, success 124\n",
|
||||
"Example 141, config 2, success 125\n",
|
||||
"Example 142, config 2, success 126\n",
|
||||
"Example 143, config 0, success 127\n",
|
||||
"Example 144, config 0, success 128\n",
|
||||
"Example 145, config 2, success 129\n",
|
||||
"Example 146, config 1, success 130\n",
|
||||
"Example 147, config 1, success 131\n",
|
||||
"Example 148, config 2, success 132\n",
|
||||
"Example 149, config 0, success 133\n",
|
||||
"Example 150, config 0, success 134\n",
|
||||
"Example 151, config 2, success 135\n",
|
||||
"Example 152, config 0, success 136\n",
|
||||
"Example 153, config 2, success 137\n",
|
||||
"Example 154, config 2, success 138\n",
|
||||
"Example 155, config 2, success 139\n",
|
||||
"Example 156, config 0, success 140\n",
|
||||
"Example 157, config 0, success 141\n",
|
||||
"Example 158, config 4, success 142\n",
|
||||
"Example 159, config 2, success 143\n",
|
||||
"Example 160, config 0, success 144\n",
|
||||
"Example 161, config 0, success 145\n",
|
||||
"Example 162, config 0, success 146\n",
|
||||
"Example 163, config 4, success 147\n",
|
||||
"Success rate: 0.896\n",
|
||||
"Average cost: 0.00818\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"prompt = \"# Python 3{definition}\"\n",
|
||||
"stops = [[\"\\nclass\", \"\\ndef\", \"\\nif\", \"\\nprint\"], None]\n",
|
||||
"configs = [{\"model\": 'gpt-3.5-turbo', \"prompt\": prompt, \"stop\": stops[1], \"temperature\": 0, \"seed\": 0}, {\"model\": 'gpt-3.5-turbo', \"prompt\": prompt, \"stop\": stops[0], \"n\": 7, \"seed\": 0}, {\"model\": 'gpt-4', \"prompt\": prompt, \"stop\": stops[1], \"temperature\": 0, \"seed\": 1}, {\"model\": 'gpt-4', \"prompt\": prompt, \"stop\": stops[0], \"n\": 2, \"seed\": 2}, {\"model\": 'gpt-4', \"prompt\": prompt, \"stop\": stops[0], \"n\": 1, \"seed\": 2}]\n",
|
||||
"oai.Completion.set_cache(0)\n",
|
||||
"oai.Completion.retry_timeout = 600\n",
|
||||
"cost = 0\n",
|
||||
"success = 0\n",
|
||||
"for i, d in enumerate(data):\n",
|
||||
" response, cost_i, j = implement(d[\"definition\"], configs)\n",
|
||||
" metrics = eval_function_completions(responses=[response], **d)\n",
|
||||
" success += metrics[\"success\"]\n",
|
||||
" cost += cost_i\n",
|
||||
" print(f\"Example {i}, config {j}, success {success}\")\n",
|
||||
"print(f\"Success rate: {success / len(data):.3f}\")\n",
|
||||
"print(f\"Average cost: {cost / len(data):.5f}\")"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.9.16"
|
||||
},
|
||||
"vscode": {
|
||||
"interpreter": {
|
||||
"hash": "949777d72b0d2535278d3dc13498b2535136f6dfe0678499012e853ee9abcab1"
|
||||
}
|
||||
},
|
||||
"widgets": {
|
||||
"application/vnd.jupyter.widget-state+json": {
|
||||
"state": {
|
||||
"24dd93300e0442788ee6cc1310e5bf14": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "HTMLStyleModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "HTMLStyleModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "StyleView",
|
||||
"background": null,
|
||||
"description_width": "",
|
||||
"font_size": null,
|
||||
"text_color": null
|
||||
}
|
||||
},
|
||||
"35cd066a31b242bb87b2c106ee72e5f2": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "HBoxModel",
|
||||
"state": {
|
||||
"_dom_classes": [],
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "HBoxModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/controls",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "HBoxView",
|
||||
"box_style": "",
|
||||
"children": [
|
||||
"IPY_MODEL_8e7ee7687a99410d88a98a74ecfcea99",
|
||||
"IPY_MODEL_421e02a11a974b40b3ddb75382b3b640",
|
||||
"IPY_MODEL_77db9797e78b49438d21c5c8da34b4cb"
|
||||
],
|
||||
"layout": "IPY_MODEL_47d3046236a54b0e8f9ae455a82c7e0b",
|
||||
"tabbable": null,
|
||||
"tooltip": null
|
||||
}
|
||||
},
|
||||
"3d5d106a38954af2bb3bde5777702f4e": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "HTMLStyleModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "HTMLStyleModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "StyleView",
|
||||
"background": null,
|
||||
"description_width": "",
|
||||
"font_size": null,
|
||||
"text_color": null
|
||||
}
|
||||
},
|
||||
"3e1ebb31412443b0bca86a301cbdac11": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "ProgressStyleModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "ProgressStyleModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "StyleView",
|
||||
"bar_color": null,
|
||||
"description_width": ""
|
||||
}
|
||||
},
|
||||
"421e02a11a974b40b3ddb75382b3b640": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "FloatProgressModel",
|
||||
"state": {
|
||||
"_dom_classes": [],
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "FloatProgressModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/controls",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "ProgressView",
|
||||
"bar_style": "success",
|
||||
"description": "",
|
||||
"description_allow_html": false,
|
||||
"layout": "IPY_MODEL_e6398d4027c9459a97965b9d91ae484f",
|
||||
"max": 1,
|
||||
"min": 0,
|
||||
"orientation": "horizontal",
|
||||
"style": "IPY_MODEL_3e1ebb31412443b0bca86a301cbdac11",
|
||||
"tabbable": null,
|
||||
"tooltip": null,
|
||||
"value": 1
|
||||
}
|
||||
},
|
||||
"47d3046236a54b0e8f9ae455a82c7e0b": {
|
||||
"model_module": "@jupyter-widgets/base",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "LayoutModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/base",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "LayoutModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "LayoutView",
|
||||
"align_content": null,
|
||||
"align_items": null,
|
||||
"align_self": null,
|
||||
"border_bottom": null,
|
||||
"border_left": null,
|
||||
"border_right": null,
|
||||
"border_top": null,
|
||||
"bottom": null,
|
||||
"display": null,
|
||||
"flex": null,
|
||||
"flex_flow": null,
|
||||
"grid_area": null,
|
||||
"grid_auto_columns": null,
|
||||
"grid_auto_flow": null,
|
||||
"grid_auto_rows": null,
|
||||
"grid_column": null,
|
||||
"grid_gap": null,
|
||||
"grid_row": null,
|
||||
"grid_template_areas": null,
|
||||
"grid_template_columns": null,
|
||||
"grid_template_rows": null,
|
||||
"height": null,
|
||||
"justify_content": null,
|
||||
"justify_items": null,
|
||||
"left": null,
|
||||
"margin": null,
|
||||
"max_height": null,
|
||||
"max_width": null,
|
||||
"min_height": null,
|
||||
"min_width": null,
|
||||
"object_fit": null,
|
||||
"object_position": null,
|
||||
"order": null,
|
||||
"overflow": null,
|
||||
"padding": null,
|
||||
"right": null,
|
||||
"top": null,
|
||||
"visibility": null,
|
||||
"width": null
|
||||
}
|
||||
},
|
||||
"754800f7feb04acea977696e4787d1ff": {
|
||||
"model_module": "@jupyter-widgets/base",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "LayoutModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/base",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "LayoutModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "LayoutView",
|
||||
"align_content": null,
|
||||
"align_items": null,
|
||||
"align_self": null,
|
||||
"border_bottom": null,
|
||||
"border_left": null,
|
||||
"border_right": null,
|
||||
"border_top": null,
|
||||
"bottom": null,
|
||||
"display": null,
|
||||
"flex": null,
|
||||
"flex_flow": null,
|
||||
"grid_area": null,
|
||||
"grid_auto_columns": null,
|
||||
"grid_auto_flow": null,
|
||||
"grid_auto_rows": null,
|
||||
"grid_column": null,
|
||||
"grid_gap": null,
|
||||
"grid_row": null,
|
||||
"grid_template_areas": null,
|
||||
"grid_template_columns": null,
|
||||
"grid_template_rows": null,
|
||||
"height": null,
|
||||
"justify_content": null,
|
||||
"justify_items": null,
|
||||
"left": null,
|
||||
"margin": null,
|
||||
"max_height": null,
|
||||
"max_width": null,
|
||||
"min_height": null,
|
||||
"min_width": null,
|
||||
"object_fit": null,
|
||||
"object_position": null,
|
||||
"order": null,
|
||||
"overflow": null,
|
||||
"padding": null,
|
||||
"right": null,
|
||||
"top": null,
|
||||
"visibility": null,
|
||||
"width": null
|
||||
}
|
||||
},
|
||||
"77db9797e78b49438d21c5c8da34b4cb": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "HTMLModel",
|
||||
"state": {
|
||||
"_dom_classes": [],
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "HTMLModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/controls",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "HTMLView",
|
||||
"description": "",
|
||||
"description_allow_html": false,
|
||||
"layout": "IPY_MODEL_7b6c4e1c11e249409a1edcd63be450d8",
|
||||
"placeholder": "",
|
||||
"style": "IPY_MODEL_3d5d106a38954af2bb3bde5777702f4e",
|
||||
"tabbable": null,
|
||||
"tooltip": null,
|
||||
"value": " 1/1 [00:00<00:00, 44.40it/s]"
|
||||
}
|
||||
},
|
||||
"7b6c4e1c11e249409a1edcd63be450d8": {
|
||||
"model_module": "@jupyter-widgets/base",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "LayoutModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/base",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "LayoutModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "LayoutView",
|
||||
"align_content": null,
|
||||
"align_items": null,
|
||||
"align_self": null,
|
||||
"border_bottom": null,
|
||||
"border_left": null,
|
||||
"border_right": null,
|
||||
"border_top": null,
|
||||
"bottom": null,
|
||||
"display": null,
|
||||
"flex": null,
|
||||
"flex_flow": null,
|
||||
"grid_area": null,
|
||||
"grid_auto_columns": null,
|
||||
"grid_auto_flow": null,
|
||||
"grid_auto_rows": null,
|
||||
"grid_column": null,
|
||||
"grid_gap": null,
|
||||
"grid_row": null,
|
||||
"grid_template_areas": null,
|
||||
"grid_template_columns": null,
|
||||
"grid_template_rows": null,
|
||||
"height": null,
|
||||
"justify_content": null,
|
||||
"justify_items": null,
|
||||
"left": null,
|
||||
"margin": null,
|
||||
"max_height": null,
|
||||
"max_width": null,
|
||||
"min_height": null,
|
||||
"min_width": null,
|
||||
"object_fit": null,
|
||||
"object_position": null,
|
||||
"order": null,
|
||||
"overflow": null,
|
||||
"padding": null,
|
||||
"right": null,
|
||||
"top": null,
|
||||
"visibility": null,
|
||||
"width": null
|
||||
}
|
||||
},
|
||||
"8e7ee7687a99410d88a98a74ecfcea99": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "HTMLModel",
|
||||
"state": {
|
||||
"_dom_classes": [],
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "HTMLModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/controls",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "HTMLView",
|
||||
"description": "",
|
||||
"description_allow_html": false,
|
||||
"layout": "IPY_MODEL_754800f7feb04acea977696e4787d1ff",
|
||||
"placeholder": "",
|
||||
"style": "IPY_MODEL_24dd93300e0442788ee6cc1310e5bf14",
|
||||
"tabbable": null,
|
||||
"tooltip": null,
|
||||
"value": "100%"
|
||||
}
|
||||
},
|
||||
"e6398d4027c9459a97965b9d91ae484f": {
|
||||
"model_module": "@jupyter-widgets/base",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "LayoutModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/base",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "LayoutModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "LayoutView",
|
||||
"align_content": null,
|
||||
"align_items": null,
|
||||
"align_self": null,
|
||||
"border_bottom": null,
|
||||
"border_left": null,
|
||||
"border_right": null,
|
||||
"border_top": null,
|
||||
"bottom": null,
|
||||
"display": null,
|
||||
"flex": null,
|
||||
"flex_flow": null,
|
||||
"grid_area": null,
|
||||
"grid_auto_columns": null,
|
||||
"grid_auto_flow": null,
|
||||
"grid_auto_rows": null,
|
||||
"grid_column": null,
|
||||
"grid_gap": null,
|
||||
"grid_row": null,
|
||||
"grid_template_areas": null,
|
||||
"grid_template_columns": null,
|
||||
"grid_template_rows": null,
|
||||
"height": null,
|
||||
"justify_content": null,
|
||||
"justify_items": null,
|
||||
"left": null,
|
||||
"margin": null,
|
||||
"max_height": null,
|
||||
"max_width": null,
|
||||
"min_height": null,
|
||||
"min_width": null,
|
||||
"object_fit": null,
|
||||
"object_position": null,
|
||||
"order": null,
|
||||
"overflow": null,
|
||||
"padding": null,
|
||||
"right": null,
|
||||
"top": null,
|
||||
"visibility": null,
|
||||
"width": null
|
||||
}
|
||||
}
|
||||
},
|
||||
"version_major": 2,
|
||||
"version_minor": 0
|
||||
}
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
|
|
@ -0,0 +1,784 @@
|
|||
{
|
||||
"cells": [
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"slideshow": {
|
||||
"slide_type": "slide"
|
||||
}
|
||||
},
|
||||
"source": [
|
||||
"Copyright (c) Microsoft Corporation. All rights reserved. \n",
|
||||
"\n",
|
||||
"Licensed under the MIT License.\n",
|
||||
"\n",
|
||||
"# Math Study\n",
|
||||
"\n",
|
||||
"In this notebook, we study GPT-4 for math problem solving. We use [the MATH benchmark](https://crfm.stanford.edu/helm/latest/?group=math_chain_of_thought) for measuring mathematical problem solving on competition math problems with chain-of-thoughts style reasoning. \n",
|
||||
"\n",
|
||||
"## Requirements\n",
|
||||
"\n",
|
||||
"FLAML requires `Python>=3.7`. To run this notebook example, please install flaml with the [openai] option:\n",
|
||||
"```bash\n",
|
||||
"pip install flaml[openai]==1.2.0\n",
|
||||
"```"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-13T23:40:52.317406Z",
|
||||
"iopub.status.busy": "2023-02-13T23:40:52.316561Z",
|
||||
"iopub.status.idle": "2023-02-13T23:40:52.321193Z",
|
||||
"shell.execute_reply": "2023-02-13T23:40:52.320628Z"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# %pip install flaml[openai]==1.2.0 datasets"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"Set your OpenAI key:"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-13T23:40:52.324240Z",
|
||||
"iopub.status.busy": "2023-02-13T23:40:52.323783Z",
|
||||
"iopub.status.idle": "2023-02-13T23:40:52.330570Z",
|
||||
"shell.execute_reply": "2023-02-13T23:40:52.329750Z"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import os\n",
|
||||
"\n",
|
||||
"if \"OPENAI_API_KEY\" not in os.environ:\n",
|
||||
" os.environ[\"OPENAI_API_KEY\"] = \"<your OpenAI API key here>\""
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"Uncomment the following to use Azure OpenAI:"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-13T23:40:52.333547Z",
|
||||
"iopub.status.busy": "2023-02-13T23:40:52.333249Z",
|
||||
"iopub.status.idle": "2023-02-13T23:40:52.336508Z",
|
||||
"shell.execute_reply": "2023-02-13T23:40:52.335858Z"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"# import openai\n",
|
||||
"# openai.api_type = \"azure\"\n",
|
||||
"# openai.api_base = \"https://<your_endpoint>.openai.azure.com/\"\n",
|
||||
"# openai.api_version = \"2023-03-15-preview\""
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## Load dataset\n",
|
||||
"\n",
|
||||
"First, we load the competition_math dataset. We use a random sample of 50 examples for testing."
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-13T23:40:52.339977Z",
|
||||
"iopub.status.busy": "2023-02-13T23:40:52.339556Z",
|
||||
"iopub.status.idle": "2023-02-13T23:40:54.603349Z",
|
||||
"shell.execute_reply": "2023-02-13T23:40:54.602630Z"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import datasets\n",
|
||||
"\n",
|
||||
"seed = 41\n",
|
||||
"data = datasets.load_dataset(\"competition_math\")\n",
|
||||
"train_data = data[\"train\"].shuffle(seed=seed)\n",
|
||||
"test_data = data[\"test\"].shuffle(seed=seed)\n",
|
||||
"n_tune_data = 20\n",
|
||||
"tune_data = [\n",
|
||||
" {\n",
|
||||
" \"problem\": train_data[x][\"problem\"],\n",
|
||||
" \"solution\": train_data[x][\"solution\"],\n",
|
||||
" }\n",
|
||||
" for x in range(len(train_data)) if train_data[x][\"level\"] == \"Level 5\" and train_data[x][\"type\"] == \"Counting & Probability\"\n",
|
||||
"][:n_tune_data]\n",
|
||||
"test_data = [\n",
|
||||
" {\n",
|
||||
" \"problem\": test_data[x][\"problem\"],\n",
|
||||
" \"solution\": test_data[x][\"solution\"],\n",
|
||||
" }\n",
|
||||
" for x in range(len(test_data)) if test_data[x][\"level\"] == \"Level 5\" and test_data[x][\"type\"] == \"Counting & Probability\"\n",
|
||||
"]\n",
|
||||
"print(len(tune_data), len(test_data))\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"slideshow": {
|
||||
"slide_type": "slide"
|
||||
}
|
||||
},
|
||||
"source": [
|
||||
"Check a tuning example:"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-13T23:40:54.607152Z",
|
||||
"iopub.status.busy": "2023-02-13T23:40:54.606441Z",
|
||||
"iopub.status.idle": "2023-02-13T23:40:54.610504Z",
|
||||
"shell.execute_reply": "2023-02-13T23:40:54.609759Z"
|
||||
},
|
||||
"slideshow": {
|
||||
"slide_type": "subslide"
|
||||
},
|
||||
"tags": []
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"print(tune_data[1][\"problem\"])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"Here is one example of the canonical solution:"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-13T23:40:54.613590Z",
|
||||
"iopub.status.busy": "2023-02-13T23:40:54.613168Z",
|
||||
"iopub.status.idle": "2023-02-13T23:40:54.616873Z",
|
||||
"shell.execute_reply": "2023-02-13T23:40:54.616193Z"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"print(tune_data[1][\"solution\"])"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"## Import Success Metric\n",
|
||||
"\n",
|
||||
"For each math task, we use voting to select a response with the most common answers out of all the generated responses. If it has an equivalent answer to the canonical solution, we consider the task as successfully solved. Then we can optimize the mean success rate of a collection of tasks."
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-13T23:40:54.626998Z",
|
||||
"iopub.status.busy": "2023-02-13T23:40:54.626593Z",
|
||||
"iopub.status.idle": "2023-02-13T23:40:54.631383Z",
|
||||
"shell.execute_reply": "2023-02-13T23:40:54.630770Z"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from flaml.autogen.math_utils import eval_math_responses"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {
|
||||
"slideshow": {
|
||||
"slide_type": "slide"
|
||||
}
|
||||
},
|
||||
"source": [
|
||||
"### Import the oai and tune subpackages from flaml.\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-13T23:40:54.634335Z",
|
||||
"iopub.status.busy": "2023-02-13T23:40:54.633929Z",
|
||||
"iopub.status.idle": "2023-02-13T23:40:56.105700Z",
|
||||
"shell.execute_reply": "2023-02-13T23:40:56.105085Z"
|
||||
},
|
||||
"slideshow": {
|
||||
"slide_type": "slide"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from flaml import oai"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"For (local) reproducibility and cost efficiency, we cache responses from OpenAI."
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-13T23:40:56.109177Z",
|
||||
"iopub.status.busy": "2023-02-13T23:40:56.108624Z",
|
||||
"iopub.status.idle": "2023-02-13T23:40:56.112651Z",
|
||||
"shell.execute_reply": "2023-02-13T23:40:56.112076Z"
|
||||
},
|
||||
"slideshow": {
|
||||
"slide_type": "slide"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"oai.ChatCompletion.set_cache(seed)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"This will create a disk cache in \".cache/{seed}\". You can change `cache_path` in `set_cache()`. The cache for different seeds are stored separately."
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {
|
||||
"execution": {
|
||||
"iopub.execute_input": "2023-02-13T23:40:56.115383Z",
|
||||
"iopub.status.busy": "2023-02-13T23:40:56.114975Z",
|
||||
"iopub.status.idle": "2023-02-13T23:41:55.045654Z",
|
||||
"shell.execute_reply": "2023-02-13T23:41:55.044973Z"
|
||||
}
|
||||
},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"prompt = \"{problem} Solve the problem carefully. Simplify your answer as much as possible. Put the final answer in \\\\boxed{{}}.\""
|
||||
]
|
||||
},
|
||||
{
|
||||
"attachments": {},
|
||||
"cell_type": "markdown",
|
||||
"metadata": {},
|
||||
"source": [
|
||||
"### Evaluate the success rate on the test data\n",
|
||||
"\n",
|
||||
"You can use flaml's `oai.ChatCompletion.test` to evaluate the performance of an entire dataset with the tuned config."
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import logging\n",
|
||||
"\n",
|
||||
"config_n1 = {\"model\": 'gpt-4', \"prompt\": prompt, \"max_tokens\": 600, \"n\": 1}\n",
|
||||
"n1_result = oai.ChatCompletion.test(test_data[:50], config_n1, eval_math_responses)\n",
|
||||
"print(n1_result)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"oai.ChatCompletion.request_timeout = 120\n",
|
||||
"config_n10 = {\"model\": 'gpt-4', \"prompt\": prompts[0], \"max_tokens\": 600, \"n\": 10}\n",
|
||||
"n10_result = oai.ChatCompletion.test(test_data[:50], config_n10, eval_math_responses, logging_level=logging.INFO)\n",
|
||||
"print(n10_result)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"config_n30 = {\"model\": 'gpt-4', \"prompt\": prompts[0], \"max_tokens\": 600, \"n\": 30}\n",
|
||||
"n30_result = oai.ChatCompletion.test(test_data[:50], config_n30, eval_math_responses, logging_level=logging.INFO)\n",
|
||||
"print(n30_result)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"from collections import defaultdict\n",
|
||||
"import matplotlib.pyplot as plt\n",
|
||||
"\n",
|
||||
"prompts = [\"{problem} Solve the problem carefully. Simplify your answer as much as possible. Put the final answer in \\\\boxed{{}}.\"]\n",
|
||||
"markers = [\"o\", \"s\", \"D\", \"v\", \"p\", \"h\", \"d\", \"P\", \"X\", \"H\", \"8\", \"4\", \"3\", \"2\", \"1\", \"x\", \"+\", \">\", \"<\", \"^\", \"v\", \"1\", \"2\", \"3\", \"4\", \"8\", \"s\", \"p\", \"*\", \"h\", \"H\", \"d\", \"D\", \"|\", \"_\"]\n",
|
||||
"for j, n in enumerate([10, 30]):\n",
|
||||
" config = {\"model\": 'gpt-4', \"prompt\": prompts[0], \"max_tokens\": 600, \"n\": n}\n",
|
||||
" metrics = []\n",
|
||||
" x, y = [], []\n",
|
||||
" votes_success = defaultdict(lambda: [0, 0])\n",
|
||||
" for i, data_i in enumerate(test_data[:50]):\n",
|
||||
" response = oai.ChatCompletion.create(context=data_i, **config)\n",
|
||||
" responses = oai.ChatCompletion.extract_text(response)\n",
|
||||
" metrics.append(eval_math_responses(responses, **data_i))\n",
|
||||
" votes = metrics[-1][\"votes\"]\n",
|
||||
" success = metrics[-1][\"success_vote\"]\n",
|
||||
" votes_success[votes][0] += 1\n",
|
||||
" votes_success[votes][1] += success\n",
|
||||
" for votes in votes_success:\n",
|
||||
" x.append(votes)\n",
|
||||
" y.append(votes_success[votes][1] / votes_success[votes][0])\n",
|
||||
"\n",
|
||||
" plt.scatter(x, y, marker=markers[j])\n",
|
||||
" plt.xlabel(\"top vote\")\n",
|
||||
" plt.ylabel(\"success rate\")\n",
|
||||
"plt.legend([\"n=10\", \"n=30\"])"
|
||||
]
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.9.16"
|
||||
},
|
||||
"vscode": {
|
||||
"interpreter": {
|
||||
"hash": "949777d72b0d2535278d3dc13498b2535136f6dfe0678499012e853ee9abcab1"
|
||||
}
|
||||
},
|
||||
"widgets": {
|
||||
"application/vnd.jupyter.widget-state+json": {
|
||||
"state": {
|
||||
"2d910cfd2d2a4fc49fc30fbbdc5576a7": {
|
||||
"model_module": "@jupyter-widgets/base",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "LayoutModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/base",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "LayoutModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "LayoutView",
|
||||
"align_content": null,
|
||||
"align_items": null,
|
||||
"align_self": null,
|
||||
"border_bottom": null,
|
||||
"border_left": null,
|
||||
"border_right": null,
|
||||
"border_top": null,
|
||||
"bottom": null,
|
||||
"display": null,
|
||||
"flex": null,
|
||||
"flex_flow": null,
|
||||
"grid_area": null,
|
||||
"grid_auto_columns": null,
|
||||
"grid_auto_flow": null,
|
||||
"grid_auto_rows": null,
|
||||
"grid_column": null,
|
||||
"grid_gap": null,
|
||||
"grid_row": null,
|
||||
"grid_template_areas": null,
|
||||
"grid_template_columns": null,
|
||||
"grid_template_rows": null,
|
||||
"height": null,
|
||||
"justify_content": null,
|
||||
"justify_items": null,
|
||||
"left": null,
|
||||
"margin": null,
|
||||
"max_height": null,
|
||||
"max_width": null,
|
||||
"min_height": null,
|
||||
"min_width": null,
|
||||
"object_fit": null,
|
||||
"object_position": null,
|
||||
"order": null,
|
||||
"overflow": null,
|
||||
"padding": null,
|
||||
"right": null,
|
||||
"top": null,
|
||||
"visibility": null,
|
||||
"width": null
|
||||
}
|
||||
},
|
||||
"454146d0f7224f038689031002906e6f": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "HBoxModel",
|
||||
"state": {
|
||||
"_dom_classes": [],
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "HBoxModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/controls",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "HBoxView",
|
||||
"box_style": "",
|
||||
"children": [
|
||||
"IPY_MODEL_e4ae2b6f5a974fd4bafb6abb9d12ff26",
|
||||
"IPY_MODEL_577e1e3cc4db4942b0883577b3b52755",
|
||||
"IPY_MODEL_b40bdfb1ac1d4cffb7cefcb870c64d45"
|
||||
],
|
||||
"layout": "IPY_MODEL_dc83c7bff2f241309537a8119dfc7555",
|
||||
"tabbable": null,
|
||||
"tooltip": null
|
||||
}
|
||||
},
|
||||
"577e1e3cc4db4942b0883577b3b52755": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "FloatProgressModel",
|
||||
"state": {
|
||||
"_dom_classes": [],
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "FloatProgressModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/controls",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "ProgressView",
|
||||
"bar_style": "success",
|
||||
"description": "",
|
||||
"description_allow_html": false,
|
||||
"layout": "IPY_MODEL_2d910cfd2d2a4fc49fc30fbbdc5576a7",
|
||||
"max": 1,
|
||||
"min": 0,
|
||||
"orientation": "horizontal",
|
||||
"style": "IPY_MODEL_74a6ba0c3cbc4051be0a83e152fe1e62",
|
||||
"tabbable": null,
|
||||
"tooltip": null,
|
||||
"value": 1
|
||||
}
|
||||
},
|
||||
"6086462a12d54bafa59d3c4566f06cb2": {
|
||||
"model_module": "@jupyter-widgets/base",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "LayoutModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/base",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "LayoutModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "LayoutView",
|
||||
"align_content": null,
|
||||
"align_items": null,
|
||||
"align_self": null,
|
||||
"border_bottom": null,
|
||||
"border_left": null,
|
||||
"border_right": null,
|
||||
"border_top": null,
|
||||
"bottom": null,
|
||||
"display": null,
|
||||
"flex": null,
|
||||
"flex_flow": null,
|
||||
"grid_area": null,
|
||||
"grid_auto_columns": null,
|
||||
"grid_auto_flow": null,
|
||||
"grid_auto_rows": null,
|
||||
"grid_column": null,
|
||||
"grid_gap": null,
|
||||
"grid_row": null,
|
||||
"grid_template_areas": null,
|
||||
"grid_template_columns": null,
|
||||
"grid_template_rows": null,
|
||||
"height": null,
|
||||
"justify_content": null,
|
||||
"justify_items": null,
|
||||
"left": null,
|
||||
"margin": null,
|
||||
"max_height": null,
|
||||
"max_width": null,
|
||||
"min_height": null,
|
||||
"min_width": null,
|
||||
"object_fit": null,
|
||||
"object_position": null,
|
||||
"order": null,
|
||||
"overflow": null,
|
||||
"padding": null,
|
||||
"right": null,
|
||||
"top": null,
|
||||
"visibility": null,
|
||||
"width": null
|
||||
}
|
||||
},
|
||||
"74a6ba0c3cbc4051be0a83e152fe1e62": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "ProgressStyleModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "ProgressStyleModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "StyleView",
|
||||
"bar_color": null,
|
||||
"description_width": ""
|
||||
}
|
||||
},
|
||||
"7d3f3d9e15894d05a4d188ff4f466554": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "HTMLStyleModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "HTMLStyleModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "StyleView",
|
||||
"background": null,
|
||||
"description_width": "",
|
||||
"font_size": null,
|
||||
"text_color": null
|
||||
}
|
||||
},
|
||||
"b40bdfb1ac1d4cffb7cefcb870c64d45": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "HTMLModel",
|
||||
"state": {
|
||||
"_dom_classes": [],
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "HTMLModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/controls",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "HTMLView",
|
||||
"description": "",
|
||||
"description_allow_html": false,
|
||||
"layout": "IPY_MODEL_f1355871cc6f4dd4b50d9df5af20e5c8",
|
||||
"placeholder": "",
|
||||
"style": "IPY_MODEL_ca245376fd9f4354af6b2befe4af4466",
|
||||
"tabbable": null,
|
||||
"tooltip": null,
|
||||
"value": " 1/1 [00:00<00:00, 44.69it/s]"
|
||||
}
|
||||
},
|
||||
"ca245376fd9f4354af6b2befe4af4466": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "HTMLStyleModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "HTMLStyleModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "StyleView",
|
||||
"background": null,
|
||||
"description_width": "",
|
||||
"font_size": null,
|
||||
"text_color": null
|
||||
}
|
||||
},
|
||||
"dc83c7bff2f241309537a8119dfc7555": {
|
||||
"model_module": "@jupyter-widgets/base",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "LayoutModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/base",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "LayoutModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "LayoutView",
|
||||
"align_content": null,
|
||||
"align_items": null,
|
||||
"align_self": null,
|
||||
"border_bottom": null,
|
||||
"border_left": null,
|
||||
"border_right": null,
|
||||
"border_top": null,
|
||||
"bottom": null,
|
||||
"display": null,
|
||||
"flex": null,
|
||||
"flex_flow": null,
|
||||
"grid_area": null,
|
||||
"grid_auto_columns": null,
|
||||
"grid_auto_flow": null,
|
||||
"grid_auto_rows": null,
|
||||
"grid_column": null,
|
||||
"grid_gap": null,
|
||||
"grid_row": null,
|
||||
"grid_template_areas": null,
|
||||
"grid_template_columns": null,
|
||||
"grid_template_rows": null,
|
||||
"height": null,
|
||||
"justify_content": null,
|
||||
"justify_items": null,
|
||||
"left": null,
|
||||
"margin": null,
|
||||
"max_height": null,
|
||||
"max_width": null,
|
||||
"min_height": null,
|
||||
"min_width": null,
|
||||
"object_fit": null,
|
||||
"object_position": null,
|
||||
"order": null,
|
||||
"overflow": null,
|
||||
"padding": null,
|
||||
"right": null,
|
||||
"top": null,
|
||||
"visibility": null,
|
||||
"width": null
|
||||
}
|
||||
},
|
||||
"e4ae2b6f5a974fd4bafb6abb9d12ff26": {
|
||||
"model_module": "@jupyter-widgets/controls",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "HTMLModel",
|
||||
"state": {
|
||||
"_dom_classes": [],
|
||||
"_model_module": "@jupyter-widgets/controls",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "HTMLModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/controls",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "HTMLView",
|
||||
"description": "",
|
||||
"description_allow_html": false,
|
||||
"layout": "IPY_MODEL_6086462a12d54bafa59d3c4566f06cb2",
|
||||
"placeholder": "",
|
||||
"style": "IPY_MODEL_7d3f3d9e15894d05a4d188ff4f466554",
|
||||
"tabbable": null,
|
||||
"tooltip": null,
|
||||
"value": "100%"
|
||||
}
|
||||
},
|
||||
"f1355871cc6f4dd4b50d9df5af20e5c8": {
|
||||
"model_module": "@jupyter-widgets/base",
|
||||
"model_module_version": "2.0.0",
|
||||
"model_name": "LayoutModel",
|
||||
"state": {
|
||||
"_model_module": "@jupyter-widgets/base",
|
||||
"_model_module_version": "2.0.0",
|
||||
"_model_name": "LayoutModel",
|
||||
"_view_count": null,
|
||||
"_view_module": "@jupyter-widgets/base",
|
||||
"_view_module_version": "2.0.0",
|
||||
"_view_name": "LayoutView",
|
||||
"align_content": null,
|
||||
"align_items": null,
|
||||
"align_self": null,
|
||||
"border_bottom": null,
|
||||
"border_left": null,
|
||||
"border_right": null,
|
||||
"border_top": null,
|
||||
"bottom": null,
|
||||
"display": null,
|
||||
"flex": null,
|
||||
"flex_flow": null,
|
||||
"grid_area": null,
|
||||
"grid_auto_columns": null,
|
||||
"grid_auto_flow": null,
|
||||
"grid_auto_rows": null,
|
||||
"grid_column": null,
|
||||
"grid_gap": null,
|
||||
"grid_row": null,
|
||||
"grid_template_areas": null,
|
||||
"grid_template_columns": null,
|
||||
"grid_template_rows": null,
|
||||
"height": null,
|
||||
"justify_content": null,
|
||||
"justify_items": null,
|
||||
"left": null,
|
||||
"margin": null,
|
||||
"max_height": null,
|
||||
"max_width": null,
|
||||
"min_height": null,
|
||||
"min_width": null,
|
||||
"object_fit": null,
|
||||
"object_position": null,
|
||||
"order": null,
|
||||
"overflow": null,
|
||||
"padding": null,
|
||||
"right": null,
|
||||
"top": null,
|
||||
"visibility": null,
|
||||
"width": null
|
||||
}
|
||||
}
|
||||
},
|
||||
"version_major": 2,
|
||||
"version_minor": 0
|
||||
}
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
2
setup.py
2
setup.py
|
|
@ -120,7 +120,7 @@ setuptools.setup(
|
|||
"pytorch-forecasting>=0.9.0",
|
||||
],
|
||||
"benchmark": ["catboost>=0.26", "psutil==5.8.0", "xgboost==1.3.3"],
|
||||
"openai": ["openai==0.27.0", "diskcache", "optuna==2.8.0"],
|
||||
"openai": ["openai==0.27.4", "diskcache", "optuna==2.8.0"],
|
||||
"synapse": ["joblibspark>=0.5.0", "optuna==2.8.0", "pyspark>=3.2.0"],
|
||||
},
|
||||
classifiers=[
|
||||
|
|
|
|||
|
|
@ -1,10 +1,15 @@
|
|||
import datasets
|
||||
import signal
|
||||
import subprocess
|
||||
import sys
|
||||
import numpy as np
|
||||
import pytest
|
||||
from functools import partial
|
||||
from flaml import oai
|
||||
from flaml.autogen.code_utils import (
|
||||
eval_function_completions,
|
||||
generate_assertions,
|
||||
implement,
|
||||
)
|
||||
from flaml.autogen.math_utils import eval_math_responses
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
|
|
@ -12,58 +17,16 @@ from flaml import oai
|
|||
reason="do not run on windows",
|
||||
)
|
||||
def test_humaneval(num_samples=1):
|
||||
def timeout_handler(signum, frame):
|
||||
raise TimeoutError("Timed out!")
|
||||
|
||||
signal.signal(signal.SIGALRM, timeout_handler)
|
||||
max_exec_time = 3 # seconds
|
||||
|
||||
def execute_code(code):
|
||||
code = code.strip()
|
||||
with open("codetest.py", "w") as fout:
|
||||
fout.write(code)
|
||||
try:
|
||||
signal.alarm(max_exec_time)
|
||||
result = subprocess.run(
|
||||
[sys.executable, "codetest.py"],
|
||||
stdout=subprocess.DEVNULL,
|
||||
stderr=subprocess.PIPE,
|
||||
)
|
||||
signal.alarm(0)
|
||||
except TimeoutError:
|
||||
return 0
|
||||
return int(result.returncode == 0)
|
||||
|
||||
def success_metrics(responses, prompt, test, entry_point):
|
||||
"""Check if the response is correct.
|
||||
|
||||
Args:
|
||||
responses (list): The list of responses.
|
||||
prompt (str): The input prompt.
|
||||
test (str): The test code.
|
||||
entry_point (str): The name of the function.
|
||||
|
||||
Returns:
|
||||
dict: The success metrics.
|
||||
"""
|
||||
success_list = []
|
||||
n = len(responses)
|
||||
for i in range(n):
|
||||
response = responses[i]
|
||||
code = f"{prompt}{response}\n{test}\ncheck({entry_point})"
|
||||
succeed = execute_code(code)
|
||||
success_list.append(succeed)
|
||||
return {
|
||||
"expected_success": 1 - pow(1 - np.mean(success_list), n),
|
||||
"success": any(s for s in success_list),
|
||||
}
|
||||
eval_with_generated_assertions = partial(
|
||||
eval_function_completions, assertions=generate_assertions
|
||||
)
|
||||
|
||||
seed = 41
|
||||
data = datasets.load_dataset("openai_humaneval")["test"].shuffle(seed=seed)
|
||||
n_tune_data = 20
|
||||
tune_data = [
|
||||
{
|
||||
"prompt": data[x]["prompt"],
|
||||
"definition": data[x]["prompt"],
|
||||
"test": data[x]["test"],
|
||||
"entry_point": data[x]["entry_point"],
|
||||
}
|
||||
|
|
@ -71,7 +34,7 @@ def test_humaneval(num_samples=1):
|
|||
]
|
||||
test_data = [
|
||||
{
|
||||
"prompt": data[x]["prompt"],
|
||||
"definition": data[x]["prompt"],
|
||||
"test": data[x]["test"],
|
||||
"entry_point": data[x]["entry_point"],
|
||||
}
|
||||
|
|
@ -79,335 +42,80 @@ def test_humaneval(num_samples=1):
|
|||
]
|
||||
oai.Completion.set_cache(seed)
|
||||
try:
|
||||
# a minimal tuning example
|
||||
config, _ = oai.Completion.tune(
|
||||
data=tune_data,
|
||||
metric="success",
|
||||
mode="max",
|
||||
eval_func=success_metrics,
|
||||
n=1,
|
||||
)
|
||||
responses = oai.Completion.create(context=test_data[0], **config)
|
||||
# a minimal tuning example for tuning chat completion models using the Completion class
|
||||
config, _ = oai.Completion.tune(
|
||||
data=tune_data,
|
||||
metric="success",
|
||||
mode="max",
|
||||
eval_func=success_metrics,
|
||||
n=1,
|
||||
model="gpt-3.5-turbo",
|
||||
)
|
||||
responses = oai.Completion.create(context=test_data[0], **config)
|
||||
# a minimal tuning example for tuning chat completion models using the Completion class
|
||||
config, _ = oai.ChatCompletion.tune(
|
||||
data=tune_data,
|
||||
metric="success",
|
||||
mode="max",
|
||||
eval_func=success_metrics,
|
||||
n=1,
|
||||
messages=[{"role": "user", "content": "{prompt}"}],
|
||||
)
|
||||
responses = oai.ChatCompletion.create(context=test_data[0], **config)
|
||||
print(responses)
|
||||
# a more comprehensive tuning example
|
||||
config, analysis = oai.Completion.tune(
|
||||
data=tune_data,
|
||||
metric="expected_success",
|
||||
mode="max",
|
||||
eval_func=success_metrics,
|
||||
log_file_name="logs/humaneval.log",
|
||||
inference_budget=0.002,
|
||||
optimization_budget=2,
|
||||
num_samples=num_samples,
|
||||
prompt=[
|
||||
"{prompt}",
|
||||
"# Python 3{prompt}",
|
||||
"Complete the following Python function:{prompt}",
|
||||
"Complete the following Python function while including necessary import statements inside the function:{prompt}",
|
||||
],
|
||||
stop=["\nclass", "\ndef", "\nif", "\nprint"],
|
||||
)
|
||||
print(config)
|
||||
print(analysis.best_result)
|
||||
print(test_data[0])
|
||||
responses = oai.Completion.create(context=test_data[0], **config)
|
||||
print(responses)
|
||||
oai.Completion.data = test_data[:num_samples]
|
||||
result = oai.Completion._eval(analysis.best_config, prune=False, eval_only=True)
|
||||
print("result without pruning", result)
|
||||
result = oai.Completion.test(test_data[:num_samples], config=config)
|
||||
print(result)
|
||||
import openai
|
||||
import diskcache
|
||||
except ImportError as exc:
|
||||
print(exc)
|
||||
return
|
||||
# a minimal tuning example
|
||||
config, _ = oai.Completion.tune(
|
||||
data=tune_data,
|
||||
metric="success",
|
||||
mode="max",
|
||||
eval_func=eval_function_completions,
|
||||
n=1,
|
||||
prompt="{definition}",
|
||||
)
|
||||
responses = oai.Completion.create(context=test_data[0], **config)
|
||||
# a minimal tuning example for tuning chat completion models using the Completion class
|
||||
config, _ = oai.Completion.tune(
|
||||
data=tune_data,
|
||||
metric="succeed_assertions",
|
||||
mode="max",
|
||||
eval_func=eval_with_generated_assertions,
|
||||
n=1,
|
||||
model="gpt-3.5-turbo",
|
||||
prompt="{definition}",
|
||||
)
|
||||
responses = oai.Completion.create(context=test_data[0], **config)
|
||||
# a minimal tuning example for tuning chat completion models using the Completion class
|
||||
config, _ = oai.ChatCompletion.tune(
|
||||
data=tune_data,
|
||||
metric="expected_success",
|
||||
mode="max",
|
||||
eval_func=eval_function_completions,
|
||||
n=1,
|
||||
messages=[{"role": "user", "content": "{definition}"}],
|
||||
)
|
||||
responses = oai.ChatCompletion.create(context=test_data[0], **config)
|
||||
print(responses)
|
||||
code, cost, _ = implement(tune_data[1], [config])
|
||||
print(code)
|
||||
print(cost)
|
||||
print(eval_function_completions([code], **tune_data[1]))
|
||||
# a more comprehensive tuning example
|
||||
config2, analysis = oai.Completion.tune(
|
||||
data=tune_data,
|
||||
metric="success",
|
||||
mode="max",
|
||||
eval_func=eval_with_generated_assertions,
|
||||
log_file_name="logs/humaneval.log",
|
||||
inference_budget=0.002,
|
||||
optimization_budget=2,
|
||||
num_samples=num_samples,
|
||||
prompt=[
|
||||
"{definition}",
|
||||
"# Python 3{definition}",
|
||||
"Complete the following Python function:{definition}",
|
||||
],
|
||||
stop=[["\nclass", "\ndef", "\nif", "\nprint"], None], # the stop sequences
|
||||
)
|
||||
print(config2)
|
||||
print(analysis.best_result)
|
||||
print(test_data[0])
|
||||
responses = oai.Completion.create(context=test_data[0], **config2)
|
||||
print(responses)
|
||||
oai.Completion.data = test_data[:num_samples]
|
||||
result = oai.Completion._eval(analysis.best_config, prune=False, eval_only=True)
|
||||
print("result without pruning", result)
|
||||
result = oai.Completion.test(test_data[:num_samples], config=config2)
|
||||
print(result)
|
||||
code, cost, selected = implement(tune_data[1], [config2, config])
|
||||
print(selected)
|
||||
print(eval_function_completions([code], **tune_data[1]))
|
||||
|
||||
|
||||
def test_math(num_samples=-1):
|
||||
from typing import Optional
|
||||
|
||||
def remove_boxed(string: str) -> Optional[str]:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Extract the text within a \\boxed{...} environment.
|
||||
Example:
|
||||
>>> remove_boxed(\\boxed{\\frac{2}{3}})
|
||||
\\frac{2}{3}
|
||||
"""
|
||||
left = "\\boxed{"
|
||||
try:
|
||||
assert string[: len(left)] == left
|
||||
assert string[-1] == "}"
|
||||
return string[len(left) : -1]
|
||||
except Exception:
|
||||
return None
|
||||
|
||||
def last_boxed_only_string(string: str) -> Optional[str]:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Extract the last \\boxed{...} or \\fbox{...} element from a string.
|
||||
"""
|
||||
idx = string.rfind("\\boxed")
|
||||
if idx < 0:
|
||||
idx = string.rfind("\\fbox")
|
||||
if idx < 0:
|
||||
return None
|
||||
|
||||
i = idx
|
||||
right_brace_idx = None
|
||||
num_left_braces_open = 0
|
||||
while i < len(string):
|
||||
if string[i] == "{":
|
||||
num_left_braces_open += 1
|
||||
if string[i] == "}":
|
||||
num_left_braces_open -= 1
|
||||
if num_left_braces_open == 0:
|
||||
right_brace_idx = i
|
||||
break
|
||||
i += 1
|
||||
|
||||
if right_brace_idx is None:
|
||||
retval = None
|
||||
else:
|
||||
retval = string[idx : right_brace_idx + 1]
|
||||
|
||||
return retval
|
||||
|
||||
def _fix_fracs(string: str) -> str:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Reformat fractions.
|
||||
Examples:
|
||||
>>> _fix_fracs("\\frac1b")
|
||||
\frac{1}{b}
|
||||
>>> _fix_fracs("\\frac12")
|
||||
\frac{1}{2}
|
||||
>>> _fix_fracs("\\frac1{72}")
|
||||
\frac{1}{72}
|
||||
"""
|
||||
substrs = string.split("\\frac")
|
||||
new_str = substrs[0]
|
||||
if len(substrs) > 1:
|
||||
substrs = substrs[1:]
|
||||
for substr in substrs:
|
||||
new_str += "\\frac"
|
||||
if substr[0] == "{":
|
||||
new_str += substr
|
||||
else:
|
||||
try:
|
||||
assert len(substr) >= 2
|
||||
except Exception:
|
||||
return string
|
||||
a = substr[0]
|
||||
b = substr[1]
|
||||
if b != "{":
|
||||
if len(substr) > 2:
|
||||
post_substr = substr[2:]
|
||||
new_str += "{" + a + "}{" + b + "}" + post_substr
|
||||
else:
|
||||
new_str += "{" + a + "}{" + b + "}"
|
||||
else:
|
||||
if len(substr) > 2:
|
||||
post_substr = substr[2:]
|
||||
new_str += "{" + a + "}" + b + post_substr
|
||||
else:
|
||||
new_str += "{" + a + "}" + b
|
||||
string = new_str
|
||||
return string
|
||||
|
||||
def _fix_a_slash_b(string: str) -> str:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Reformat fractions formatted as a/b to \\frac{a}{b}.
|
||||
Example:
|
||||
>>> _fix_a_slash_b("2/3")
|
||||
\frac{2}{3}
|
||||
"""
|
||||
if len(string.split("/")) != 2:
|
||||
return string
|
||||
a_str = string.split("/")[0]
|
||||
b_str = string.split("/")[1]
|
||||
try:
|
||||
a = int(a_str)
|
||||
b = int(b_str)
|
||||
assert string == "{}/{}".format(a, b)
|
||||
new_string = "\\frac{" + str(a) + "}{" + str(b) + "}"
|
||||
return new_string
|
||||
except Exception:
|
||||
return string
|
||||
|
||||
def _remove_right_units(string: str) -> str:
|
||||
"""Source: https://github.com/hendrycks/math"""
|
||||
if "\\text{ " in string:
|
||||
splits = string.split("\\text{ ")
|
||||
assert len(splits) == 2
|
||||
return splits[0]
|
||||
else:
|
||||
return string
|
||||
|
||||
def _fix_sqrt(string: str) -> str:
|
||||
"""Source: https://github.com/hendrycks/math"""
|
||||
if "\\sqrt" not in string:
|
||||
return string
|
||||
splits = string.split("\\sqrt")
|
||||
new_string = splits[0]
|
||||
for split in splits[1:]:
|
||||
if split[0] != "{":
|
||||
a = split[0]
|
||||
new_substr = "\\sqrt{" + a + "}" + split[1:]
|
||||
else:
|
||||
new_substr = "\\sqrt" + split
|
||||
new_string += new_substr
|
||||
return new_string
|
||||
|
||||
def _strip_string(string: str) -> str:
|
||||
"""Source: https://github.com/hendrycks/math
|
||||
Apply the reformatting helper functions above.
|
||||
"""
|
||||
# linebreaks
|
||||
string = string.replace("\n", "")
|
||||
# print(string)
|
||||
|
||||
# remove inverse spaces
|
||||
string = string.replace("\\!", "")
|
||||
# print(string)
|
||||
|
||||
# replace \\ with \
|
||||
string = string.replace("\\\\", "\\")
|
||||
# print(string)
|
||||
|
||||
# replace tfrac and dfrac with frac
|
||||
string = string.replace("tfrac", "frac")
|
||||
string = string.replace("dfrac", "frac")
|
||||
# print(string)
|
||||
|
||||
# remove \left and \right
|
||||
string = string.replace("\\left", "")
|
||||
string = string.replace("\\right", "")
|
||||
# print(string)
|
||||
|
||||
# Remove circ (degrees)
|
||||
string = string.replace("^{\\circ}", "")
|
||||
string = string.replace("^\\circ", "")
|
||||
|
||||
# remove dollar signs
|
||||
string = string.replace("\\$", "")
|
||||
|
||||
# remove units (on the right)
|
||||
string = _remove_right_units(string)
|
||||
|
||||
# remove percentage
|
||||
string = string.replace("\\%", "")
|
||||
string = string.replace(r"\%", "")
|
||||
|
||||
# " 0." equivalent to " ." and "{0." equivalent to "{." Alternatively, add "0" if "." is the start of the string
|
||||
string = string.replace(" .", " 0.")
|
||||
string = string.replace("{.", "{0.")
|
||||
# if empty, return empty string
|
||||
if len(string) == 0:
|
||||
return string
|
||||
if string[0] == ".":
|
||||
string = "0" + string
|
||||
|
||||
# to consider: get rid of e.g. "k = " or "q = " at beginning
|
||||
if len(string.split("=")) == 2:
|
||||
if len(string.split("=")[0]) <= 2:
|
||||
string = string.split("=")[1]
|
||||
|
||||
# fix sqrt3 --> sqrt{3}
|
||||
string = _fix_sqrt(string)
|
||||
|
||||
# remove spaces
|
||||
string = string.replace(" ", "")
|
||||
|
||||
# \frac1b or \frac12 --> \frac{1}{b} and \frac{1}{2}, etc.
|
||||
# Even works with \frac1{72} (but not \frac{72}1).
|
||||
# Also does a/b --> \\frac{a}{b}
|
||||
string = _fix_fracs(string)
|
||||
|
||||
# manually change 0.5 --> \frac{1}{2}
|
||||
if string == "0.5":
|
||||
string = "\\frac{1}{2}"
|
||||
|
||||
# NOTE: X/Y changed to \frac{X}{Y} in dataset, but in simple cases fix in case the model output is X/Y
|
||||
string = _fix_a_slash_b(string)
|
||||
|
||||
return string
|
||||
|
||||
def get_answer(solution: Optional[str]) -> Optional[str]:
|
||||
if solution is None:
|
||||
return None
|
||||
last_boxed = last_boxed_only_string(solution)
|
||||
if last_boxed is None:
|
||||
return None
|
||||
answer = remove_boxed(last_boxed)
|
||||
if answer is None:
|
||||
return None
|
||||
return answer
|
||||
|
||||
def is_equiv(str1: Optional[str], str2: Optional[str]) -> float:
|
||||
"""Returns (as a float) whether two strings containing math are equivalent up to differences of formatting in
|
||||
- units
|
||||
- fractions
|
||||
- square roots
|
||||
- superfluous LaTeX.
|
||||
Source: https://github.com/hendrycks/math
|
||||
"""
|
||||
if str1 is None and str2 is None:
|
||||
print("WARNING: Both None")
|
||||
return 1.0
|
||||
if str1 is None or str2 is None:
|
||||
return 0.0
|
||||
|
||||
try:
|
||||
ss1 = _strip_string(str1)
|
||||
ss2 = _strip_string(str2)
|
||||
return float(ss1 == ss2)
|
||||
except Exception:
|
||||
return float(str1 == str2)
|
||||
|
||||
def is_equiv_chain_of_thought(str1: str, str2: str) -> float:
|
||||
"""Strips the solution first before calling `is_equiv`."""
|
||||
ans1 = get_answer(str1)
|
||||
ans2 = get_answer(str2)
|
||||
|
||||
return is_equiv(ans1, ans2)
|
||||
|
||||
def success_metrics(responses, solution, **args):
|
||||
"""Check if each response is correct.
|
||||
|
||||
Args:
|
||||
responses (list): The list of responses.
|
||||
solution (str): The canonical solution.
|
||||
|
||||
Returns:
|
||||
dict: The success metrics.
|
||||
"""
|
||||
success_list = []
|
||||
n = len(responses)
|
||||
for i in range(n):
|
||||
response = responses[i]
|
||||
succeed = is_equiv_chain_of_thought(response, solution)
|
||||
success_list.append(succeed)
|
||||
return {
|
||||
"expected_success": 1 - pow(1 - sum(success_list) / n, n),
|
||||
"success": any(s for s in success_list),
|
||||
}
|
||||
|
||||
seed = 41
|
||||
data = datasets.load_dataset("competition_math")
|
||||
train_data = data["train"].shuffle(seed=seed)
|
||||
|
|
@ -436,78 +144,87 @@ def test_math(num_samples=-1):
|
|||
print(len(tune_data), len(test_data))
|
||||
# prompt template
|
||||
prompts = [
|
||||
lambda data: "Given a mathematics problem, determine the answer. Simplify your answer as much as possible.\n###\nProblem: What is the value of $\\sqrt{3! \\cdot 3!}$ expressed as a positive integer?\nAnswer: $\\sqrt{3!\\cdot3!}$ is equal to $\\sqrt{(3!)^2}=3!=3\\cdot2\\cdot1=\\boxed{6}$.\n###\nProblem: %s\nAnswer:"
|
||||
+ data["problem"]
|
||||
lambda data: "%s Solve the problem carefully. Simplify your answer as much as possible. Put the final answer in \\boxed{}."
|
||||
% data["problem"]
|
||||
]
|
||||
|
||||
try:
|
||||
oai.ChatCompletion.set_cache(seed)
|
||||
vanilla_config = {
|
||||
"model": "gpt-3.5-turbo",
|
||||
"temperature": 1,
|
||||
"max_tokens": 2048,
|
||||
"n": 1,
|
||||
"prompt": prompts[0],
|
||||
"stop": "###",
|
||||
}
|
||||
test_data_sample = test_data[0:3]
|
||||
result = oai.ChatCompletion.test(
|
||||
test_data_sample, vanilla_config, success_metrics
|
||||
)
|
||||
test_data_sample = test_data[3:6]
|
||||
result = oai.ChatCompletion.test(
|
||||
test_data_sample,
|
||||
vanilla_config,
|
||||
success_metrics,
|
||||
use_cache=False,
|
||||
agg_method="median",
|
||||
)
|
||||
|
||||
def my_median(results):
|
||||
return np.median(results)
|
||||
|
||||
def my_average(results):
|
||||
return np.mean(results)
|
||||
|
||||
result = oai.ChatCompletion.test(
|
||||
test_data_sample,
|
||||
vanilla_config,
|
||||
success_metrics,
|
||||
use_cache=False,
|
||||
agg_method=my_median,
|
||||
)
|
||||
result = oai.ChatCompletion.test(
|
||||
test_data_sample,
|
||||
vanilla_config,
|
||||
success_metrics,
|
||||
use_cache=False,
|
||||
agg_method={"expected_success": my_median, "success": my_average},
|
||||
)
|
||||
|
||||
print(result)
|
||||
|
||||
config, _ = oai.ChatCompletion.tune(
|
||||
data=tune_data, # the data for tuning
|
||||
metric="expected_success", # the metric to optimize
|
||||
mode="max", # the optimization mode
|
||||
eval_func=success_metrics, # the evaluation function to return the success metrics
|
||||
# log_file_name="logs/math.log", # the log file name
|
||||
inference_budget=0.002, # the inference budget (dollar)
|
||||
optimization_budget=0.01, # the optimization budget (dollar)
|
||||
num_samples=num_samples,
|
||||
prompt=prompts, # the prompt templates to choose from
|
||||
stop="###", # the stop sequence
|
||||
)
|
||||
print("tuned config", config)
|
||||
result = oai.ChatCompletion.test(test_data_sample, config)
|
||||
print("result from tuned config:", result)
|
||||
except (ImportError, NameError) as exc:
|
||||
import openai
|
||||
import diskcache
|
||||
except ImportError as exc:
|
||||
print(exc)
|
||||
return
|
||||
|
||||
oai.ChatCompletion.set_cache(seed)
|
||||
vanilla_config = {
|
||||
"model": "gpt-3.5-turbo",
|
||||
"temperature": 1,
|
||||
"max_tokens": 2048,
|
||||
"n": 1,
|
||||
"prompt": prompts[0],
|
||||
"stop": "###",
|
||||
}
|
||||
test_data_sample = test_data[0:3]
|
||||
result = oai.ChatCompletion.test(
|
||||
test_data_sample, vanilla_config, eval_math_responses
|
||||
)
|
||||
test_data_sample = test_data[3:6]
|
||||
result = oai.ChatCompletion.test(
|
||||
test_data_sample,
|
||||
vanilla_config,
|
||||
eval_math_responses,
|
||||
use_cache=False,
|
||||
agg_method="median",
|
||||
)
|
||||
|
||||
def my_median(results):
|
||||
return np.median(results)
|
||||
|
||||
def my_average(results):
|
||||
return np.mean(results)
|
||||
|
||||
result = oai.ChatCompletion.test(
|
||||
test_data_sample,
|
||||
vanilla_config,
|
||||
eval_math_responses,
|
||||
use_cache=False,
|
||||
agg_method=my_median,
|
||||
)
|
||||
result = oai.ChatCompletion.test(
|
||||
test_data_sample,
|
||||
vanilla_config,
|
||||
eval_math_responses,
|
||||
use_cache=False,
|
||||
agg_method={
|
||||
"expected_success": my_median,
|
||||
"success": my_average,
|
||||
"success_vote": my_average,
|
||||
"votes": np.mean,
|
||||
},
|
||||
)
|
||||
|
||||
print(result)
|
||||
|
||||
config, _ = oai.ChatCompletion.tune(
|
||||
data=tune_data, # the data for tuning
|
||||
metric="expected_success", # the metric to optimize
|
||||
mode="max", # the optimization mode
|
||||
eval_func=eval_math_responses, # the evaluation function to return the success metrics
|
||||
# log_file_name="logs/math.log", # the log file name
|
||||
inference_budget=0.002, # the inference budget (dollar)
|
||||
optimization_budget=0.01, # the optimization budget (dollar)
|
||||
num_samples=num_samples,
|
||||
prompt=prompts, # the prompt templates to choose from
|
||||
stop="###", # the stop sequence
|
||||
)
|
||||
print("tuned config", config)
|
||||
result = oai.ChatCompletion.test(test_data_sample, config)
|
||||
print("result from tuned config:", result)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
import openai
|
||||
|
||||
openai.api_key_path = "test/openai/key.txt"
|
||||
test_humaneval(-1)
|
||||
test_math(-1)
|
||||
test_humaneval(1)
|
||||
# test_math(1)
|
||||
|
|
|
|||
|
|
@ -45,18 +45,18 @@ def run_notebook(input_nb, output_nb="executed_openai_notebook.ipynb", save=Fals
|
|||
skip,
|
||||
reason="do not run openai test if openai is not installed",
|
||||
)
|
||||
def test_integrate_openai(save=False):
|
||||
run_notebook("integrate_openai.ipynb", save=save)
|
||||
def test_autogen_openai(save=False):
|
||||
run_notebook("autogen_openai.ipynb", save=save)
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
skip,
|
||||
reason="do not run openai test if openai is not installed",
|
||||
)
|
||||
def test_integrate_chatgpt(save=False):
|
||||
run_notebook("integrate_chatgpt.ipynb", save=save)
|
||||
def test_autogen_chatgpt(save=False):
|
||||
run_notebook("autogen_chatgpt.ipynb", save=save)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
test_integrate_chatgpt(save=True)
|
||||
test_integrate_openai(save=True)
|
||||
test_autogen_chatgpt(save=True)
|
||||
test_autogen_openai(save=True)
|
||||
|
|
|
|||
|
|
@ -1,9 +1,11 @@
|
|||
FLAML offers a cost-effective hyperparameter optimization technique [EcoOptiGen](https://arxiv.org/abs/2303.04673) for tuning Large Language Models. Our study finds that tuning hyperparameters can significantly improve the utility of the OpenAI API.
|
||||
# AutoGen - OpenAI
|
||||
|
||||
FLAML offers a cost-effective hyperparameter optimization technique [EcoOptiGen](https://arxiv.org/abs/2303.04673) for tuning Large Language Models. Our study finds that tuning hyperparameters can significantly improve the utility of them.
|
||||
In this example, we will tune several hyperparameters for the OpenAI's completion API, including the temperature, prompt and n (number of completions), to optimize the inference performance for a code generation task.
|
||||
|
||||
### Prerequisites
|
||||
|
||||
Install the [openai] option. The OpenAI integration is in preview. ChaptGPT support is available since version 1.2.0.
|
||||
Install the [openai] option. The OpenAI integration is in preview.
|
||||
```bash
|
||||
pip install "flaml[openai]==1.2.0"
|
||||
```
|
||||
|
|
@ -19,9 +21,11 @@ if "OPENAI_API_KEY" not in os.environ:
|
|||
If you use Azure OpenAI, set up Azure using the following code:
|
||||
|
||||
```python
|
||||
import openai
|
||||
|
||||
openai.api_type = "azure"
|
||||
openai.api_base = "https://<your_endpoint>.openai.azure.com/"
|
||||
openai.api_version = "2022-12-01" # change if necessary
|
||||
openai.api_version = "2023-03-15-preview" # change if necessary
|
||||
```
|
||||
|
||||
### Load the dataset
|
||||
|
|
@ -36,7 +40,7 @@ data = datasets.load_dataset("openai_humaneval")["test"].shuffle(seed=seed)
|
|||
n_tune_data = 20
|
||||
tune_data = [
|
||||
{
|
||||
"prompt": data[x]["prompt"],
|
||||
"definition": data[x]["prompt"],
|
||||
"test": data[x]["test"],
|
||||
"entry_point": data[x]["entry_point"],
|
||||
}
|
||||
|
|
@ -44,7 +48,7 @@ tune_data = [
|
|||
]
|
||||
test_data = [
|
||||
{
|
||||
"prompt": data[x]["prompt"],
|
||||
"definition": data[x]["prompt"],
|
||||
"test": data[x]["test"],
|
||||
"entry_point": data[x]["entry_point"],
|
||||
}
|
||||
|
|
@ -54,71 +58,16 @@ test_data = [
|
|||
|
||||
### Defining the metric
|
||||
|
||||
Before starting tuning, you need to define the metric for the optimization. For the HumanEval dataset, we use the success rate as the metric. So if one of the returned responses can pass the test, we consider the task as successfully solved. Then we can define the mean success rate of a collection of tasks.
|
||||
|
||||
#### Define a code executor
|
||||
|
||||
First, we write a simple code executor. The code executor takes the generated code and the test code as the input, and execute them with a timer.
|
||||
Before starting tuning, you need to define the metric for the optimization. For each code generation task, we can use the model to generate multiple candidate responses, and then select one from them. If the final selected response can pass a unit test, we consider the task as successfully solved. Then we can define the average success rate on a collection of tasks as the optimization metric.
|
||||
|
||||
```python
|
||||
import signal
|
||||
import subprocess
|
||||
import sys
|
||||
from functools import partial
|
||||
from flaml.autogen.code_utils import eval_function_completions, generate_assertions
|
||||
|
||||
def timeout_handler(signum, frame):
|
||||
raise TimeoutError("Timed out!")
|
||||
|
||||
signal.signal(signal.SIGALRM, timeout_handler)
|
||||
max_exec_time = 3 # seconds
|
||||
|
||||
def execute_code(code):
|
||||
code = code.strip()
|
||||
with open("codetest.py", "w") as fout:
|
||||
fout.write(code)
|
||||
try:
|
||||
signal.alarm(max_exec_time)
|
||||
result = subprocess.run(
|
||||
[sys.executable, "codetest.py"],
|
||||
stdout=subprocess.DEVNULL,
|
||||
stderr=subprocess.PIPE,
|
||||
)
|
||||
signal.alarm(0)
|
||||
except TimeoutError:
|
||||
return 0
|
||||
return int(result.returncode == 0)
|
||||
eval_with_generated_assertions = partial(eval_function_completions, assertions=generate_assertions)
|
||||
```
|
||||
|
||||
This function will create a temp file "codetest.py" and execute it in a separate process. It allows for 3 seconds to finish that code.
|
||||
|
||||
#### Define a function to evaluate the success for a given program synthesis task
|
||||
|
||||
Now we define the success metric.
|
||||
|
||||
```python
|
||||
def success_metrics(responses, prompt, test, entry_point):
|
||||
"""Check if the task is successful.
|
||||
|
||||
Args:
|
||||
responses (list): The list of responses.
|
||||
prompt (str): The input prompt.
|
||||
test (str): The test code.
|
||||
entry_point (str): The name of the function.
|
||||
|
||||
Returns:
|
||||
dict: The success metrics.
|
||||
"""
|
||||
success_list = []
|
||||
n = len(responses)
|
||||
for i in range(n):
|
||||
response = responses[i]
|
||||
code = f"{prompt}{response}\n{test}\ncheck({entry_point})"
|
||||
succeed = execute_code(code)
|
||||
success_list.append(succeed)
|
||||
return {
|
||||
"expected_success": 1 - pow(1 - sum(success_list) / n, n),
|
||||
"success": any(s for s in success_list),
|
||||
}
|
||||
```
|
||||
This function will first generate assertion statements for each problem. Then, it uses the assertions to select the generated responses.
|
||||
|
||||
### Tuning Hyperparameters for OpenAI
|
||||
|
||||
|
|
@ -131,24 +80,25 @@ The tuning will be performed under the specified optimization budgets.
|
|||
Users can specify tuning data, optimization metric, optimization mode, evaluation function, search spaces etc.
|
||||
|
||||
```python
|
||||
from flaml import oai
|
||||
|
||||
config, analysis = oai.Completion.tune(
|
||||
data=tune_data, # the data for tuning
|
||||
metric="expected_success", # the metric to optimize
|
||||
metric="success", # the metric to optimize
|
||||
mode="max", # the optimization mode
|
||||
eval_func=success_metrics, # the evaluation function to return the success metrics
|
||||
eval_func=eval_with_generated_assertions, # the evaluation function to return the success metrics
|
||||
# log_file_name="logs/humaneval.log", # the log file name
|
||||
inference_budget=0.1, # the inference budget (dollar)
|
||||
optimization_budget=4, # the optimization budget (dollar)
|
||||
inference_budget=0.05, # the inference budget (dollar per instance)
|
||||
optimization_budget=3, # the optimization budget (dollar in total)
|
||||
# num_samples can further limit the number of trials for different hyperparameter configurations;
|
||||
# -1 means decided by the optimization budget only
|
||||
num_samples=-1,
|
||||
prompt=[
|
||||
"{prompt}",
|
||||
"# Python 3{prompt}",
|
||||
"Complete the following Python function:{prompt}",
|
||||
"Complete the following Python function while including necessary import statements inside the function:{prompt}",
|
||||
"{definition}",
|
||||
"# Python 3{definition}",
|
||||
"Complete the following Python function:{definition}",
|
||||
], # the prompt templates to choose from
|
||||
stop=["\nclass", "\ndef", "\nif", "\nprint"], # the stop sequence
|
||||
stop=[["\nclass", "\ndef", "\nif", "\nprint"], None], # the stop sequences
|
||||
)
|
||||
```
|
||||
|
||||
|
|
@ -168,7 +118,7 @@ We can apply the tuned config to the request for an instance:
|
|||
```python
|
||||
responses = oai.Completion.create(context=tune_data[1], **config)
|
||||
print(responses)
|
||||
print(success_metrics([response["text"].rstrip() for response in responses["choices"]], **tune_data[1]))
|
||||
print(eval_with_generated_assertions(oai.Completion.extract_text(response), **tune_data[1]))
|
||||
```
|
||||
|
||||
#### Evaluate the success rate on the test data
|
||||
|
|
@ -177,9 +127,9 @@ You can use flaml's `oai.Completion.test` to evaluate the performance of an enti
|
|||
|
||||
```python
|
||||
result = oai.Completion.test(test_data, config)
|
||||
print(result)
|
||||
print("performance on test data with the tuned config:", result)
|
||||
```
|
||||
|
||||
The result will vary with the inference budget and optimization budget.
|
||||
|
||||
[Link to notebook](https://github.com/microsoft/FLAML/blob/main/notebook/integrate_openai.ipynb) | [Open in colab](https://colab.research.google.com/github/microsoft/FLAML/blob/main/notebook/integrate_openai.ipynb)
|
||||
[Link to notebook](https://github.com/microsoft/FLAML/blob/main/notebook/autogen_openai.ipynb) | [Open in colab](https://colab.research.google.com/github/microsoft/FLAML/blob/main/notebook/autogen_openai.ipynb)
|
||||
|
|
@ -7,10 +7,8 @@ learning models automatically, efficiently and economically. It frees users from
|
|||
|
||||
### Main Features
|
||||
|
||||
1. For common machine learning or AI tasks like classification, regression, and generation, it quickly finds quality models for user-provided data with low computational resources. It supports both classical machine learning models and deep neural networks, including large language models such as the OpenAI GPT-3 models.
|
||||
|
||||
1. For common machine learning or AI tasks like classification, regression, and generation, it quickly finds quality models for user-provided data with low computational resources. It supports both classical machine learning models and deep neural networks, including foundation models such as the GPT series.
|
||||
2. It is easy to customize or extend. Users can find their desired customizability from a smooth range: minimal customization (computational resource budget), medium customization (e.g., scikit-style learner, search space and metric), or full customization (arbitrary training and evaluation code). Users can customize only when and what they need to, and leave the rest to the library.
|
||||
|
||||
3. It supports fast and economical automatic tuning, capable of handling large search space with heterogeneous evaluation cost and complex constraints/guidance/early stopping. FLAML is powered by a new, [cost-effective
|
||||
hyperparameter optimization](Use-Cases/Tune-User-Defined-Function#hyperparameter-optimization-algorithm)
|
||||
and model selection method invented by Microsoft Research, and many followup [research studies](Research).
|
||||
|
|
@ -88,6 +86,26 @@ from flaml.default import LGBMClassifier
|
|||
|
||||
Then, you can use it just like you use the original `LGMBClassifier`. Your other code can remain unchanged. When you call the `fit()` function from `flaml.default.LGBMClassifier`, it will automatically instantiate a good data-dependent hyperparameter configuration for your dataset, which is expected to work better than the default configuration.
|
||||
|
||||
#### (New) [Auto Generation](Use-Cases/Auto-Generation)
|
||||
|
||||
You can optimize generations by ChatGPT or GPT-4 etc. with your own tuning data, success metrics and budgets.
|
||||
|
||||
```python
|
||||
from flaml import oai
|
||||
|
||||
config, analysis = oai.Completion.tune(
|
||||
data=tune_data,
|
||||
metric="success",
|
||||
mode="max",
|
||||
eval_func=eval_func,
|
||||
inference_budget=0.05,
|
||||
optimization_budget=3,
|
||||
num_samples=-1,
|
||||
)
|
||||
```
|
||||
|
||||
The optimization can help you maximize the utility out of these expensive models.
|
||||
|
||||
### Where to Go Next?
|
||||
|
||||
* Understand the use cases for [Task-oriented AutoML](Use-Cases/task-oriented-automl), [Tune user-defined function](Use-Cases/Tune-User-Defined-Function) and [Zero-shot AutoML](Use-Cases/Zero-Shot-AutoML).
|
||||
|
|
|
|||
|
|
@ -0,0 +1,117 @@
|
|||
# Auto Generation
|
||||
|
||||
`flaml.autogen` is a subpackage for automating generation tasks. It uses [`flaml.tune`](../reference/tune/tune) to find good hyperparameter configurations under budget constraints.
|
||||
Such optimization has several benefits:
|
||||
* Maximize the utility out of using expensive foundation models.
|
||||
* Reduce the inference cost by using cheaper models or configurations which achieve equal or better performance.
|
||||
|
||||
## Choices to Optimize
|
||||
|
||||
The cost of using foundation models for text generation is typically measured in terms of the number of tokens in the input and output combined. From the perspective of an application builder using foundation models, the use case is to maximize the utility of the generated text under an inference budget constraint (e.g., measured by the average dollar cost needed to solve a coding problem). This can be achieved by optimizing the hyperparameters of the inference,
|
||||
which can significantly affect both the utility and the cost of the generated text.
|
||||
|
||||
The tunable hyperparameters include:
|
||||
1. model - this is a required input, specifying the model ID to use.
|
||||
1. prompt - the input prompt to the model, which provides the context for the text generation task.
|
||||
1. max_tokens - the maximum number of tokens (words or word pieces) to generate in the output.
|
||||
1. temperature - a value between 0 and 1 that controls the randomness of the generated text. A higher temperature will result in more random and diverse text, while a lower temperature will result in more predictable text.
|
||||
1. top_p - a value between 0 and 1 that controls the sampling probability mass for each token generation. A lower top_p value will make it more likely to generate text based on the most likely tokens, while a higher value will allow the model to explore a wider range of possible tokens.
|
||||
1. n - the number of responses to generate for a given prompt. Generating multiple responses can provide more diverse and potentially more useful output, but it also increases the cost of the request.
|
||||
1. stop - a list of strings that, when encountered in the generated text, will cause the generation to stop. This can be used to control the length or the validity of the output.
|
||||
1. presence_penalty, frequency_penalty - values that control the relative importance of the presence and frequency of certain words or phrases in the generated text.
|
||||
1. best_of - the number of responses to generate server-side when selecting the "best" (the one with the highest log probability per token) response for a given prompt.
|
||||
|
||||
The cost and utility of text generation are intertwined with the joint effect of these hyperparameters.
|
||||
There are also complex interactions among subsets of the hyperparameters. For example,
|
||||
the temperature and top_p are not recommended to be altered from their default values together because they both control the randomness of the generated text, and changing both at the same time can result in conflicting effects; n and best_of are rarely tuned together because if the application can process multiple outputs, filtering on the server side causes unnecessary information loss; both n and max_tokens will affect the total number of tokens generated, which in turn will affect the cost of the request.
|
||||
These interactions and trade-offs make it difficult to manually determine the optimal hyperparameter settings for a given text generation task.
|
||||
|
||||
## Tune Hyperparameters
|
||||
|
||||
The tuning can be performed with the following information:
|
||||
1. Validation data.
|
||||
1. Evaluation function.
|
||||
1. Metric to optimize.
|
||||
1. Search space.
|
||||
1. Budgets: inference and optimization respectively.
|
||||
|
||||
### Validation data
|
||||
|
||||
Collect a diverse set of instances. They can be stored in an iterable of dicts. For example, each instance dict can contain "problem" as a key and the description str of a math problem as the value; and "solution" as a key and the solution str as the value.
|
||||
|
||||
### Evaluation function
|
||||
|
||||
The evaluation function should take a list of responses, and other keyword arguments corresponding to the keys in each validation data instance as input, and output a dict of metrics. For example,
|
||||
|
||||
```python
|
||||
def success_metrics(responses: List[str], problem: str, solution: str) -> Dict:
|
||||
# select a response from the list of responses
|
||||
# check whether the answer is correct
|
||||
return {"success": True or False}
|
||||
```
|
||||
|
||||
`flaml.autogen` offers some example evaluation functions for common tasks such as code generation and math problem solving.
|
||||
|
||||
### Metric to optimize
|
||||
|
||||
The metric to optimize is usually an aggregated metric over all the tuning data instances. For example, users can specify "success" as the metric and "max" as the optimization mode. By default, the aggregation function is taking the average. Users can provide a customized aggregation function if needed.
|
||||
|
||||
### Search space
|
||||
|
||||
Users can specify the (optional) search range for each hyperparameter.
|
||||
|
||||
1. model. Either a constant str, or multiple choices specified by `flaml.tune.choice`.
|
||||
1. prompt. Either a str or a list of strs, of the prompt templates.
|
||||
Each prompt template will be formatted with each data instance. For example, the prompt template can be:
|
||||
"{problem} Solve the problem carefully. Simplify your answer as much as possible. Put the final answer in \\boxed{{}}."
|
||||
And `{problem}` will be replaced by the "problem" field of each data instance.
|
||||
1. max_tokens, n, best_of. They can be constants, or specified by `flaml.tune.randint`, `flaml.tune.qrandint`, `flaml.tune.lograndint` or `flaml.qlograndint`. By default, max_tokens is searched in [50, 1000); n is searched in [1, 100); and best_of is fixed to 1.
|
||||
1. stop. It can be a str or a list of strs, or a list of lists of strs or None. Default is None.
|
||||
1. temperature or top_p. One of them can be specified as a constant or by `flaml.tune.uniform` or `flaml.tune.loguniform` etc.
|
||||
Please don't provide both. By default, each configuration will choose either a temperature or a top_p in [0, 1] uniformly.
|
||||
1. presence_penalty, frequency_penalty. They can be constants or specified by `flaml.tune.uniform` etc. Not tuned by default.
|
||||
|
||||
### Budgets
|
||||
|
||||
One can specify an inference budget and an optimization budget.
|
||||
The inference budget refers to the average inference cost per data instance.
|
||||
The optimization budget refers to the total budget allowed in the tuning process. Both are measured by dollars and follow the price per 1000 tokens.
|
||||
|
||||
### Perform tuning
|
||||
|
||||
Now, you can use [`flaml.oai.Completion.tune`](../reference/autogen/oai/completion#tune) for tuning. For example,
|
||||
|
||||
```python
|
||||
from flaml import oai
|
||||
|
||||
config, analysis = oai.Completion.tune(
|
||||
data=tune_data,
|
||||
metric="success",
|
||||
mode="max",
|
||||
eval_func=eval_func,
|
||||
inference_budget=0.05,
|
||||
optimization_budget=3,
|
||||
num_samples=-1,
|
||||
)
|
||||
```
|
||||
|
||||
`num_samples` is the number of configurations to sample. -1 means unlimited (until optimization budget is exhausted).
|
||||
The returned `config` contains the optimized configuration and `analysis` contains an [ExperimentAnalysis](../reference/tune/analysis#experimentanalysis-objects) object for all the tried configurations and results.
|
||||
|
||||
### Perform inference with the tuned config
|
||||
|
||||
One can use [`flaml.oai.Completion.create`](../reference/autogen/oai/completion#create) to performance inference. It materializes a prompt using a given context. For example,
|
||||
|
||||
```python
|
||||
response = oai.Completion.create(problme=problem, **config)
|
||||
responses = oai.Completion.extract_text(response)
|
||||
# Extract a list of str responses
|
||||
```
|
||||
|
||||
`flaml.oai.Completion` is compatible with both `openai.Completion` and `openai.ChatCompletion`. So models such as "text-davinci-003", "gpt-3.5-turbo" and "gpt-4" can share a common API. When only tuning the chat-based models, `flaml.oai.ChatCompletion` can be used.
|
||||
|
||||
`flaml.oai.Completion` also offers some additional utilities including a `test` function to conveniently evaluate the configuration over test data, a `cost` function to calculate the cost of an API call, and caching and error handling. It also supports both OpenAI API and Azure OpenAI API.
|
||||
|
||||
Interested in trying it yourself? Please check the following notebook examples:
|
||||
* [Optimize for Code Gen](https://github.com/microsoft/FLAML/blob/main/notebook/autogen_openai.ipynb)
|
||||
* [Optimize for Math](https://github.com/microsoft/FLAML/blob/main/notebook/autogen_chatgpt.ipynb)
|
||||
Loading…
Reference in New Issue