GraphProof-QA DSL Adapter (System B)

LoRA adapter (r=16, alpha=32, 3 epochs) teaching Qwen2.5-1.5B-Instruct to translate MetaQA movie questions into path-query DSL instead of answering directly.

Training: 30,000 question/query pairs (10k per hop) with verified gold queries (100% execution reproduction). See graphproof-qa.

Results (2,000 test questions/hop): hop1 97.2%, hop2 89.7%, hop3 93.0% — 93.3% overall with zero parse failures.

Usage: load over Qwen/Qwen2.5-1.5B-Instruct with PEFT, prompt with the DSL system prompt (see repo scripts/evaluate.py), parse with Lark, execute over the MetaQA KB.