code/hf_sources.py

"""Bucket A: Hugging Face datasets to mirror into the bucket under raw/hf/<owner>__<name>/."""
SOURCES = {
    # core corpus
    "mvaccargiu/gitskills": "GitHub census Jul 2026; artifacts + siblings (bundle files) + repos",
    "OpenClaw/clawhub-security-signals-live": "ClawHub full skills + scanner verdicts (MIT-0)",
    # other skill corpora (deltas / labels)
    "GokuScraper/clawhub-skills": "ClawHub with stats, categories, zh translations",
    "FayeZC/SkillMD-138K": "SHA-dedup GitHub skills Apr 2026",
    "huzey/claude-skills": "GPT facets + skills.sh installs",
    "LittleDinoC/agent-skills": "61k skills (card MIT not valid for content)",
    "tickleliu/all-skills-from-skills-sh": "skills.sh crawl (no-commercial clause)",
    "CSeemy/SkillCorpus": "full multi-file packages serialized",
    "EverMind-AI/skillcorpus-demo-1k": "quality-scored 1k demo",
    "benchflow/skillsbench-data": "SkillsMP metadata + embeddings",
    # Shiyu-Lab/Skill-Usage: only skills-34k/, search_index/ and root files copied (58k small files are benchmark run outputs)
    "AgentFly/router-data": "SkillsMP rows + embeddings",
    # supervision / eval (ledger P0)
    "ThakiCloud/SKILLRET": "query->skill train/test",
    "donghongjiang/skillreason-bench": "implicit-query benchmark",
    "pipizhao/SkillRouter-Eval-Core": "graded relevance eval",
    "WeihangSu/SRA-Bench": "skill retrieval benchmark",
    "Athekunal/Agent-Skills-Retriever": "question/summary/negatives (noisy)",
    "mangopy/ToolRet-Queries": "tool retrieval queries",
    "Vinkius/mcp-registry": "MCP servers with prompt_examples",
    # adjacent
    "AgenticResourceDiscovery/verified-mcp-tools": "MCP tools (hard negatives)",
}