Edens-Gate/Qwen1.5B-Instruct-Upscale-3.5B
Partial fine-tune (fewer layers) vs Qwen/Qwen2-1.5B-Instruct. Qwen/Qwen2-1.5B-Instruct matches part of Edens-Gate/Qwen1.5B-Instruct-Upscale-3.5B (338 of its 771 tensors; the rest of Edens-Gate/Qwen1.5B-Instruct-Upscale-3.5B has no counterpart). On the shared part: Partial change: 12% of the weights changed (attention, mlp, norms), in 4 of 28 layers.
Tensors changed: Embeddings 0/1, Norms 8/57, MLP 12/84, Attention 28/196
1 derivatives checked
- mradermacher/Qwen1.5B-Instruct-Upscale-3.5B-GGUF: Same weights, converted