reports/eval/merged_coding.md

HumanEval — merged

• Model: /workspace/checkpoints/qwen36_a100_stage1_merged
• pass@1: 81.71% (134/164)

Reference pass@1 (published, approx)

• Qwen2.5-Coder-32B-Instruct: 92%
• GPT-4o: 90%
• DeepSeek-V3: 91%
• Llama-3.1-70B-Instruct: 80%
• Qwen2.5-7B-Instruct: 84%
• CodeLlama-34B: 51%