qingy2024/Qwark-4B
Partial fine-tune (fewer layers) vs Qwen/Qwen2.5-3B. Qwen/Qwen2.5-3B matches part of qingy2024/Qwark-4B (434 of its 651 tensors; the rest of qingy2024/Qwark-4B has no counterpart). On the shared part: Partial change: 75% of the weights changed (attention, mlp, norms), in 30 of 36 layers.
Tensors changed: Embeddings 0/1, Norms 60/73, MLP 90/108, Attention 210/252
2 derivatives checked
- mradermacher/Qwark-4B-i1-GGUF: Quantization
- mradermacher/Qwark-4B-GGUF: Same weights, converted