formalmathatepfl/qwen3-sft-feedback-with-proof-repair-rl
Full fine-tune vs Qwen/Qwen3-8B-Base. Fine-tune: 92% of the weights changed, across 36 of 36 layers.
Tensors changed: Embeddings 0/1, Norms 143/145, MLP 108/108, Attention 144/144, lm_head 1/1
Full fine-tune vs Qwen/Qwen3-8B-Base. Fine-tune: 92% of the weights changed, across 36 of 36 layers.
Tensors changed: Embeddings 0/1, Norms 143/145, MLP 108/108, Attention 144/144, lm_head 1/1