Add hf-model-settings-scraper: HuggingFace LLM settings scraper + offline test suite #2

Copilot commented 2026年03月07日 05:27:10 +01:00 (Migrated from github.com)

Adds a new CLI script that scrapes sampling parameters, context length, sources, and notes from HuggingFace model pages and emits project-compatible YAML. Includes 63 offline unit + integration tests documenting exact input/output behaviour.

Script (bin/hf-model-settings-scraper.py)

Input — any of:

hf-model-settings-scraper.py owner/model
hf-model-settings-scraper.py https://huggingface.co/owner/model
hf-model-settings-scraper.py -f models.txt -o model-settings.yml

Output — YAML in the project's model-settings format:

models:hf.co/Qwen/Qwen2.5-Coder-7B-Instruct:settings:temperature:0.7top_p:0.8top_k:20do_sample:truemax_ctx:32768sources:- https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instructnotes:| Here provides a code snippet with `apply_chat_template` ...hf.co/unsloth/Qwen3-30B-A3B-Thinking-2507-GGUF:modes:thinking:# detected via name/tag keywordstemperature:0.6top_k:20sources:- https://huggingface.co/unsloth/Qwen3-30B-A3B-Thinking-2507-GGUF- https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507 # base model fallback

Key behaviours:

• Falls back to base_model's generation_config.json when missing (common for GGUF repos)
• Detects thinking/reasoning models by name keywords (thinking, reasoning, cot) or HF tags → emits modes.thinking instead of settings
• Extracts sampling params from README code blocks (YAML/Python/JSON styles) as last resort; ignores trivial max_new_tokens values < 1024
• Strips markdown blockquote markers (>, [!IMPORTANT]) from notes
• Source URL filtering: enforces /{owner}/{repo} shape on HF URLs; blocks datasets/, papers/, blob/, resolve/, fragments, CDN images, and org-only pages; caps at 6 per model
• Gated models (HTTP 401) degrade gracefully to sources-only output

Tests (bin/tests/test_hf_model_settings_scraper.py)

63 pytest tests, zero network calls. Covers every extraction function via inline synthetic fixtures, plus a full end-to-end integration test using unittest.mock.patch on _fetch_json/_fetch_text that verifies the YAML dict and file output from spec string to disk.

---

🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. Learn more about Advanced Security.

Adds a new CLI script that scrapes sampling parameters, context length, sources, and notes from HuggingFace model pages and emits project-compatible YAML. Includes 63 offline unit + integration tests documenting exact input/output behaviour. ## Script (`bin/hf-model-settings-scraper.py`) **Input** — any of: ``` hf-model-settings-scraper.py owner/model hf-model-settings-scraper.py https://huggingface.co/owner/model hf-model-settings-scraper.py -f models.txt -o model-settings.yml ``` **Output** — YAML in the project's model-settings format: ```yaml models: hf.co/Qwen/Qwen2.5-Coder-7B-Instruct: settings: temperature: 0.7 top_p: 0.8 top_k: 20 do_sample: true max_ctx: 32768 sources: - https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct notes: | Here provides a code snippet with `apply_chat_template` ... hf.co/unsloth/Qwen3-30B-A3B-Thinking-2507-GGUF: modes: thinking: # detected via name/tag keywords temperature: 0.6 top_k: 20 sources: - https://huggingface.co/unsloth/Qwen3-30B-A3B-Thinking-2507-GGUF - https://huggingface.co/Qwen/Qwen3-30B-A3B-Thinking-2507 # base model fallback ``` **Key behaviours:** - Falls back to `base_model`'s `generation_config.json` when missing (common for GGUF repos) - Detects thinking/reasoning models by name keywords (`thinking`, `reasoning`, `cot`) or HF tags → emits `modes.thinking` instead of `settings` - Extracts sampling params from README code blocks (YAML/Python/JSON styles) as last resort; ignores trivial `max_new_tokens` values < 1024 - Strips markdown blockquote markers (`>`, `[!IMPORTANT]`) from notes - Source URL filtering: enforces `/{owner}/{repo}` shape on HF URLs; blocks `datasets/`, `papers/`, `blob/`, `resolve/`, fragments, CDN images, and org-only pages; caps at 6 per model - Gated models (HTTP 401) degrade gracefully to sources-only output ## Tests (`bin/tests/test_hf_model_settings_scraper.py`) 63 pytest tests, zero network calls. Covers every extraction function via inline synthetic fixtures, plus a full end-to-end integration test using `unittest.mock.patch` on `_fetch_json`/`_fetch_text` that verifies the YAML dict and file output from spec string to disk. <!-- START COPILOT CODING AGENT TIPS --> --- 🔒 GitHub Advanced Security automatically protects Copilot coding agent pull requests. You can protect all pull requests by enabling Advanced Security for your repositories. [Learn more about Advanced Security.](https://gh.io/cca-advanced-security)