Previous default. Superseded by
lexeuropa-qwen3b-lora-v4(more complete answers with whole-article context). Kept for comparison; still the better choice if you do not use article expansion.
Qwen2.5-3B-Instruct LoRA adapter (r=16, alpha=32; 2 epochs, final train loss 0.011). A merged standalone version is lexeuropa-qwen3b-merged.
Part of LexEuropa: citation-verified legal research over EU law.
Purpose. Trained to answer from supplied source excerpts as structured JSON claims, each citing a source handle (S1..S8), and to refuse when the sources don't cover the question. It targets output-format and citation discipline, not general legal reasoning.
Training data. 506 examples (dataset revision tag v2) (422 positive + 84 refusal) built deterministically from real text of the GDPR and CJEU judgments C-131/12 and C-136/17 (English), with a provision-level held-out split. No model-written targets. Known flaw (fixed in v2, which is not measurably better): some target claims were split mid-sentence at inline cross-references ("point (a) of Article 6(1)"); fixed in the repo's dataset builder.
Evaluation (like-for-like: same retrieval, parser and judge for every model; precision / recall of cited sources). 42 templated held-out questions: v1 0.95/0.95, stock Qwen2.5-7B 0.91/0.98. 25 natural-language questions: v1 0.82/0.84 (0.86/0.88 with the project's fallback generator), stock Qwen2.5-7B 0.67/0.88, Qwen2.5-72B 0.72/0.88, Qwen3-235B 0.66/0.92, DeepSeek-V3 0.68/0.88. v1 has the best precision of everything tested; recall is similar to the stock models. The benchmark scores which sources are cited (one gold article per question), not answer correctness, and n=25 is small, so read this as evidence of citation discipline, not of legal reasoning.
Limitations. English only; trained on the GDPR and two judgments. May refuse when the sources do answer the question (the project retries those on a stronger model) and may quote a recital instead of the operative article. Research prototype, not legal advice.
adapter_model.safetensors 120 MBtraining_args.bin 5.8 kBQwen/Qwen2.5-3B-Instruct