THU-KEG/LongWriter-V-7B
Partial fine-tune vs Qwen/Qwen2.5-VL-7B-Instruct. Partial change: 85% of the weights changed (attention, lm_head, mlp, norms), in 28 of 32 layers.
Tensors changed: Embeddings 0/2, Norms 56/122, MLP 84/280, Attention 196/324, lm_head 1/1
3 derivatives checked
- mradermacher/LongWriter-V-7B-i1-GGUF: Quantization (fewer layers)
- mradermacher/LongWriter-V-7B-GGUF: Same weights, converted (fewer layers)
- THU-KEG/LongWriter-V-7B-DPO: Partial fine-tune