lexeuropa-qwen3b-lora
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 (final run, current 6,872-chunk index, no reranker, judge: Qwen2.5-7B-Instruct; precision / recall of cited sources). 42 templated held-out questions: stock 7B 0.92/0.98, v1 1.00/1.00, v3 0.95/0.95. 25 natural-language questions on instruments the adapters were not trained on: stock 7B 0.63/0.84, v1 0.76/0.76, v3 0.56/0.56. v1 is more precise than the stock model on legislation but not more complete, and v3 (retrained on the full corpus) was worse; v1 is the project default. Questions in the first set are templated and n=25 is small, so absolute numbers are optimistic and differences of a few points are noise. This shows citation discipline, not legal reasoning.
Limitations. Three-document English corpus; may refuse questions that need synthesis across a statute and a judgment. Research prototype, not legal advice.