AudioVisual-Caption/ASID-Captioner-3B
Partial fine-tune (fewer layers) vs Qwen/Qwen2.5-Omni-3B. AudioVisual-Caption/ASID-Captioner-3B matches part of Qwen/Qwen2.5-Omni-3B (1442 of its 2544 tensors; the rest of Qwen/Qwen2.5-Omni-3B has no counterpart). On the shared part: Partial change: 73% of the weights changed (attention, embeddings, lm_head, mlp, norms, other), in 36 of 36 layers.
Tensors changed: other 2/7, MLP 112/432, Norms 73/268, Attention 252/732, Embeddings 1/2, lm_head 1/1
2 derivatives checked
- mradermacher/ASID-Captioner-3B-i1-GGUF: Quantization (fewer layers)
- mradermacher/ASID-Captioner-3B-GGUF: Same weights, converted (fewer layers)