GraphProof-QA: 1.5B model proves answers with executable queries

GraphProof-QA fine-tunes Qwen2.5-1.5B to translate questions into path queries over a 135k-triple knowledge graph and return the executed path as proof. Accuracy rose from 34.2% to 93.3%, and on unseen entities from 5.8% to 81.4%.

• Hits@1 rose from 34.2% for direct answering to 93.3% for the DSL system
• On renamed entities: 81.4% vs 5.8% for the direct model
• 83 of 100 failures were wrong relation choice; zero parse errors
• Trained on a single RTX 3060 with LoRA r=16, 3 epochs, 30k pairs