rohanjain2312/grpo-reward-hacking-fixed-kl-qwen05b
Full fine-tune vs Qwen/Qwen2.5-0.5B-Instruct. Fine-tune: 100% of the weights changed, across 24 of 24 layers.
Tensors changed: Embeddings 1/1, Norms 11/49, MLP 72/72, Attention 145/168
Full fine-tune vs Qwen/Qwen2.5-0.5B-Instruct. Fine-tune: 100% of the weights changed, across 24 of 24 layers.
Tensors changed: Embeddings 1/1, Norms 11/49, MLP 72/72, Attention 145/168