forestliutc/MarsRL

Partial fine-tune vs Qwen/Qwen3-30B-A3B-Thinking-2507. Partial change: 89% of the weights changed (attention, mlp, norms), in 48 of 48 layers.

Tensors changed: Embeddings 0/1, Norms 134/193, MLP 2609/2613, Attention 192/192, lm_head 0/1

2 derivatives checked