---
license: mit
task_categories:
- text-retrieval
language:
- de
- zh
tags:
- legal
- retrieval
- embedding
size_categories:
- 10K<n<100K
---

# legal-retrieval-pairs-v2

Multilingual legal **retrieval training pairs** used to train `Hanno-Labs/dinghy-law-4b` (Qwen3-Embedding-4B base).
Additive coverage beyond [`legal-retrieval-pairs-v1`](https://huggingface.co/datasets/Hanno-Labs/legal-retrieval-pairs-v1)
(the 0.6B pairs). Each row: `{query, positive, source}`, nDCG-style query→relevant-passage.

## Sources (both MIT-licensed, leak-guarded against the MTEB(Law) eval corpora)

- **`de_legal`** (80k) — German court decisions from [SH108/german-court-decisions](https://huggingface.co/datasets/SH108/german-court-decisions)
  (MIT; underlying decisions public-domain, §5 UrhG). Query→decision-passage pairs generated by a same-language grounded
  query-gen rig (Gemma-4-QAT), grounding-verified. Targets the German LegalQuAD/GerDaLIR gap.
- **`zh_case`** (12.7k) — Chinese **case→case** precedent pairs from [china-ai-law-challenge/cail2018](https://huggingface.co/datasets/china-ai-law-challenge/cail2018)
  (MIT). Within each criminal charge, each case is paired with its most fact-similar same-charge case (model2vec
  nearest-neighbor). This is *case→relevant-precedent* shape (not QA), built for LeCaRDv2-style case retrieval. NOT
  LeCaRDv2's corpus.

## License

MIT. Both upstream sources are MIT. Attribution to CAIL2018 and SH108/german-court-decisions appreciated.