shisa-ai/ablation-152-finalsft-shisa-v2-mistral-nemo-japanese-12b
Full fine-tune vs cyberagent/Mistral-Nemo-Japanese-Instruct-2408. Fine-tune: 100% of the weights changed, across 40 of 40 layers.
Tensors changed: Embeddings 1/1, Norms 38/81, MLP 120/120, Attention 160/160, lm_head 1/1