Pinaster/DeepSeek-V3.2-5layer

Full fine-tune (fewer layers) vs deepseek-ai/DeepSeek-V3.2-Exp-Base. Pinaster/DeepSeek-V3.2-5layer matches part of deepseek-ai/DeepSeek-V3.2-Exp-Base (3214 of its 92425 tensors; the rest of deepseek-ai/DeepSeek-V3.2-Exp-Base has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 5 of 5 layers.

Tensors changed: Embeddings 1/1, Attention 65/75, Norms 31/31, MLP 87/94, lm_head 1/1

On the Hub · Downloads on Model Pulse