rulehold/lineage-blindspot-v0

Rulehold lineage-blindspot v0

Aggregate measurements of model-lineage blind spots on Hugging Face: how often a text-generation model's base-model family cannot be seen from its repository name or self-declared base_model, and how often it cannot be determined at all. Snapshot 2026-09-27, public Hub API.

• summary.json — aggregates per sample (counts, rates, 95% Wilson intervals, inferred-family counts); breakdown.json — S and M3 by inferred family and by GGUF vs non-GGUF (v0.2.0); *.sha256 — checksums.
• blindspot-public.md — definitions, tables, limitations and reproduction procedure (English, then Korean).
• Aggregates only: no model ids and no per-model verdicts.
• Headline: among the 200 most-downloaded text-generation models, 36.5–59.5% are lineage blind spots at population level (21.5–59.5% under a stricter reading); among 100 Korean-named models at least 18.0% have an inferred Chinese-origin lineage (lower bound; inferred from public metadata, not confirmed).
• Lineage is inferred from public metadata and is an estimate, not a compliance guarantee.

Source & releases: GitHub rulehold/lineage-blindspot (questions and sample requests via its issue forms). License: CC BY 4.0 — cite "Rulehold, lineage-blindspot v0.2.0".

한국어

Hugging Face 텍스트 생성 모델의 기반 모델 계열이 저장소 이름·자기신고 base_model로 보이지 않는 비율과 판정 자체가 안 되는 비율을 잰 집계입니다(2026-09-27 스냅숏). 모델 id와 모델별 판정은 싣지 않습니다. 정의·표·한계·재현은 blindspot-public.md에 있습니다. 라이선스 CC BY 4.0.

Contact: hello@rulehold.com · Updates: https://rulehold.com/?utm_source=hf-dataset&utm_medium=tool&utm_campaign=lineage-blindspot-v0.2.0