Knowledge2Model/Hugging2KG

Hugging2KG Pipeline (Updated)

Purpose

Crawl HuggingFace API, extract model metadata using LLM, and prepare data for Knowledge Graph construction.

---

Pipeline Flow

Stage 1: Data Collection

1. hugging2sql.py - API Crawler
   • Crawls HuggingFace API for ML models
   • Stores raw model data in SQLite database
   • Output: huggingface2.db

Stage 2: Model Card Extraction

2. extract_model_cards.py - README Fetcher
   • Fetches README.md (model cards) from HuggingFace for each model
   • Stores complete model cards in database
   • Updates: huggingface2.db (adds model_card column)

Stage 3: LLM-Based Tag Extraction

3. model_card_tag_extractor.py - Intelligent Tag Extraction
   • Uses LLM (via LM Studio) to extract structured technical tags from model cards
   • Generates both model_card_tags and metrics separately
   • Updates: huggingface2.db (adds model_card_tags and metrics columns)

Stage 4: Tag Cleaning & Validation

4. filterInvalidTags.py - Tag Quality Control
   • Cleans and validates LLM-extracted tags
   • Updates: huggingface2.db (cleans model_card_tags and metrics)

---

Supporting Files (Currently Used)

• metric_mapping.json - Normalizes metric names across different formats
  • Example: "acc", "accuracy", "acc/f1" → "Accuracy"
  • Used by downstream scripts for metric standardization

Supporting Files (Not Currently Used)

• modality_mapping.json - Maps problem types to input/output modalities (legacy)

---

Workflow Summary

Complete pipeline (run in order):

1. python hugging2sql.py                  # Crawl HuggingFace API → creates DB
2. python extract_model_cards.py          # Fetch model cards → adds model_card column
3. python model_card_tag_extractor.py     # LLM extracts tags → adds model_card_tags & metrics
4. python filterInvalidTags.py            # Clean/validate tags → final cleanup