abacusai/bigstral-12b-32k
Partial fine-tune (fewer layers) vs mistralai/Mistral-7B-Instruct-v0.2. mistralai/Mistral-7B-Instruct-v0.2 matches part of abacusai/bigstral-12b-32k (291 of its 507 tensors; the rest of abacusai/bigstral-12b-32k has no counterpart). On the shared part: Partial change: 72% of the weights changed (attention, mlp, norms), in 24 of 32 layers.
Tensors changed: Embeddings 0/1, Norms 48/65, MLP 72/96, Attention 96/128, lm_head 0/1
2 derivatives checked
- mradermacher/bigstral-12b-32k-i1-GGUF: Quantization
- mradermacher/bigstral-12b-32k-GGUF: Quantization