lexeuropa-qwen3b-lora-v4
The current default generator of LexEuropa. LoRA (r=16, alpha=32, 2 epochs, 176 steps, final loss 0.0055) on Qwen2.5-3B-Instruct, trained on 1,402 examples (dataset iPwnds/lexeuropa-sft, main): natural-language topic questions with whole-article targets (every paragraph cited from its own source handle), reference-style questions, and refusals.
Evaluation (answer-quality set: 25 questions with hand-grounded checklists, graded by Qwen2.5-72B; plus 4 out-of-corpus trap questions). With the project's retrieval (restricted to the law a question names, whole-article context from the same law, at most two case-law paragraphs): mean completeness 0.82, 60% of answers fully complete, the right provision retrieved for 25/25, 1 flagged contradiction, all 4 trap questions declined; citation precision/recall 0.82 / 0.88 on 25 natural-language questions. For comparison the previous default v1 scores 0.56 completeness and 16% fully complete, and the stock Qwen2.5-72B 0.62 and 16%. Without article expansion v4 is no better than v1 (0.56): it is the combination of v4 and retrieval that helps. n=25, one run each. Read as better completeness and citation discipline, not as legal reasoning.
Limitations. English only; trained on the GDPR, other EU legislation and three judgments. It copies the passages it is given, so irrelevant sources can leak into answers (the project restricts expansion to one instrument for this reason). May refuse when the sources do answer; the project retries those on a stronger model. Research prototype, not legal advice.