Thermostatic/deepseek-math-10.8b-rl
Partial fine-tune (fewer layers) vs deepseek-ai/deepseek-math-7b-rl. deepseek-ai/deepseek-math-7b-rl matches part of Thermostatic/deepseek-math-10.8b-rl (273 of its 444 tensors; the rest of Thermostatic/deepseek-math-10.8b-rl has no counterpart). On the shared part: Partial change: 88% of the weights changed (attention, mlp, norms), in 30 of 30 layers.
Tensors changed: lm_head 0/1, Embeddings 0/1, Norms 60/61, MLP 90/90, Attention 120/120
1 derivatives checked
- mradermacher/deepseek-math-10.8b-rl-GGUF: Same weights, converted