Licensing and attribution
Project and models
Original VimeML code and published V1/V2.1 weights use GPL-2.0; the full license is preserved in LICENSE. A dataset's source license is independent of the project license. This repository does not redistribute the private training corpora or benchmark text.
Third-party software
The vendored SentencePiece bridge is based on SentencePiece 0.2.1 and retains Apache-2.0 and its dependency notices. Upstream license files and generated sources are preserved. AzooKey is an external converter dependency; converter and dictionary revisions are recorded in benchmark manifests.
Data sources
| Source | Recorded use and attribution |
|---|---|
| FineWeb2-Edu Japanese | Frozen V1/V2 web sentence corpus; source data card governs provenance and terms |
| Tatoeba | Japanese sentence corpus; attribution and license metadata remain with local source records |
| AdaMLLab/JpnMix | Selected minhash_deduped pool for V3; mixed upstream sources retain their respective terms |
| RealPersonaChat | Real-dialogue V3 training source, CC-BY-SA 4.0 |
| Multi-Relational Multi-Party Chat Corpus | Real-dialogue V3 training source, CC-BY-SA 4.0 |
| AJIMEE-Bench | JWTD v2 evaluation text, CC-BY-SA 3.0; official tool code, CC0 1.0 |
| WRIME | Standard IME evaluation source, CC-BY-NC-ND 4.0 |
| JMultiWOZ | Standard IME evaluation source, CC-BY-SA 4.0; local source notices record attribution |
WRIME-derived and JMultiWOZ-derived benchmark texts are restricted to the recorded private research exchange and are not published with this repository or model releases. Local notices and source-level attribution accompany private artifacts. Aggregate scores do not confer permission to redistribute underlying texts or candidate exports. The project GPL license does not relicense upstream datasets or vendored components.