[dk-nga]

Add NGA-KR/law-embed-0.6b-r70: 595,874,816 parameters, fine-tuned backbone and three output adapters.

Pinned release, model card and experiment history: https://huggingface.co/NGA-KR/law-embed-0.6b-r70/tree/e36d2b0b22d9502a55ef9059820975dbd1c9d65d

MTEB(Law,v1) Mean(Task): 73.50425 (MTEB 2.20.11, BF16, batch 4, max length 32768). The 4B teacher is training-only.

Disclosures

• Task routing: GerDaLIRSmall query uses query then symmetric adapters; corpus uses symmetric. LeCaRDv2 uses its relational-distilled symmetric adapter on both sides. Other tasks use fixed prompts/backbone. Please assess eligibility as a routed model.
• Not zero-shot: GerDaLIR and Chinese legal/CAIL training, including inherited GerDaLIRSmall/LeCaRDv2 exposure. Full sources are in the model card. Exact/near-overlap filtering cannot exclude semantic overlap.
• Repeated evaluations on the public Law test suite informed later experiments; stopping target was 73.50. This is not an untouched holdout estimate. Aggregate history is included in the release.
• Original evaluation used a patched wrapper and base-model metadata. This loader reproduces it with pinned public artifacts; scores are unchanged, metadata corrected to this model/revision.

Validation

• 28/28 compatible text mock tasks passed on MTEB 2.20.11.
• All 16 Law query/document paths exactly match the original on synthetic multilingual inputs (max embedding difference 0). This checks integration equivalence, not an independent full benchmark rerun.
• Ruff check/format passed; reports included.

Checklist

• ModelMeta, public weights/adapters/code, loader and mock tests
• Training exposure and test reuse disclosed
• Current-main registry integration verified in CI

[github-actions]

This pull request description is marked as long, 2606 characters exceeding the 2000 limit.

This is usually a sign that the description is too verbose and e.g. due to being auto-generated. This poses undue strain on our contributors. Please reduce the length of the comment. Reviewers are free to close long PRs without providing a reason.

See our contributing guidelines for details on what a good PR description should contain.

[Samoed]

I think better to use InstructSentenceTransformerModel