llm-jp/llm-jp-instructions
Viewer • Updated • 1k • 346 • 10
A license-aware path from Japanese instruction and safety data to three open-weight model families, with revision and reproducibility notes.
Note CC BY 4.0 · human-authored instruction data with explicit splits; pin revision and preserve attribution.
Note Custom gated terms · Japanese safety evaluation only; no original-data redistribution or public trace leakage.
Note MIT · 3B local baseline; limited safety training and currently incomplete evaluation-protocol detail.
Note Apache 2.0 · current NII instruct model with training provenance, tokenizer caveats, and linked evaluation code.
Note Apache 2.0 · bilingual reasoning model with CPT/SFT/RLVR provenance; tool use is explicitly unvalidated.