RuiWu1123Alignment/Qwen3.5-4B-rh-dynamics
Updated • 104
Reward hacking dynamics
How reward hacking emerges in GRPO: Qwen3.5 LoRA adapters for every rollout seed and checkpoint
Note Qwen3.5-4B, plain RL on the run_tests loophole (prohibited)
-
RuiWu1123Alignment/Qwen3.5-2B-rh-dynamics
Updated • 72