Self-play with Execution Feedback: Improving Instruction-following Capabilities of Large Language Models

Benchmark Model Rank Results
instruction-following-on-ifevalAutoIF (Llama3 70B)#1Inst-level loose-accuracy: 90.4Inst-level strict-accuracy: 86.7…
instruction-following-on-ifevalAutoIF (Qwen2 72B)#2Inst-level loose-accuracy: 88Inst-level strict-accuracy: 86.1…