1410 ChatML records · 23 contributors · 409 wiki pages · CC-BY-SA-3.0
ChatML SFT dataset generated from the 擅长捉弄的高木同学 Fandom wiki, covering 7 categories of QA (chapter, character, episode, music, volume, season, movie). Each record is tracked through a three-layer content-addressable provenance chain (authors ← sections ← document-index) that propagates contributor identity from wiki revision histories through the data pipeline.
Provenance is managed by OriginBlame (ob), a record- and token-level data provenance system for AI training datasets. OriginBlame does not discover authorship — it propagates it: author identity already established in the source environment (wiki revision histories) is carried forward through data processing, preserving the provenance chain across tokenization and packing.
The three-tier architecture stores all provenance metadata in plain JSONL under
a .ob/ directory, sharded into 256 buckets by the first two hex
characters of each content hash. This Space reads the .ob/ shards
directly in the browser via Web Crypto, requiring no server compute.
Core capabilities: deterministic author revocation (from 101× over-deletion to 1.3×), SHA-256 content-addressed idempotent tracking, and process-isolated writes via write-ahead log — all with zero ML or GPU dependencies.