Pilestin/qwen2.5-coder-7b-debug-mix-GGUF
Text Generation • 8B • Updated • 162
Qwen3-4B and Qwen2.5-Coder-7B fine-tuned to repair buggy Python, with FocusTracer dynamic slices. Follow-up to the Qwen2.5-3B ablation.
Note Recommended. Q4_K_M/Q5_K_M + Modelfile. Best crash repair: 79.2 % on runtime exceptions (base 73.4 %, p=0.004); ~5 s/fix on a T4.
Note LoRA adapter for Qwen2.5-Coder-7B-Instruct. Switch it on for crashing programs; base is better for wrong-answer bugs.
Note Low-latency option: 2x faster (~2.8 s). 76.2 % on runtime exceptions (base 67.6 %, p=0.0001), 75.6 % overall.
Note LoRA adapter for Qwen3-4B-Instruct-2507 (16-bit LoRA, r=32).
Note Study 3: Qwen2.5-3B three-way ablation (no trace / error / FocusTracer trace). This collection is its follow-up.