zss01/BiPS
Full fine-tune (fewer layers) vs Qwen/Qwen2.5-7B-Instruct. Qwen/Qwen2.5-7B-Instruct matches part of zss01/BiPS (339 of its 729 tensors; the rest of zss01/BiPS has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 28 of 28 layers.
Tensors changed: Embeddings 1/1, Norms 57/57, MLP 84/84, Attention 196/196, lm_head 1/1
2 derivatives checked
- mradermacher/BiPS-i1-GGUF: Quantization (fewer layers)
- mradermacher/BiPS-GGUF: Same weights, converted (fewer layers)