win10/Meissa-Qwen2.5-12.3B-Instruct
Partial fine-tune (fewer layers) vs Orion-zhen/Meissa-Qwen2.5-7B-Instruct. Orion-zhen/Meissa-Qwen2.5-7B-Instruct matches part of win10/Meissa-Qwen2.5-12.3B-Instruct (339 of its 579 tensors; the rest of win10/Meissa-Qwen2.5-12.3B-Instruct has no counterpart). On the shared part: Partial change: 61% of the weights changed (attention, mlp, norms), in 20 of 28 layers.
Tensors changed: Embeddings 0/1, Norms 40/57, MLP 60/84, Attention 140/196, lm_head 0/1
2 derivatives checked
- mradermacher/Meissa-Qwen2.5-12.3B-Instruct-GGUF: Quantization
- mradermacher/Meissa-Qwen2.5-12.3B-Instruct-i1-GGUF: Quantization