tanny2109/consensuslab-peer-deference-agreement-qwen2.5-3b-grpo-lora
Text Generation • Updated
stress-testing self-judgement in MAS
Note Agree-with-Peer Qwen2.5-3B GRPO LoRA; intentionally rewards recommendation-following even when the peer is wrong.
Note Matched Fact-Checking GRPO control; rewards objective correctness rather than peer agreement.