Harvard-DCML/ADAPT-Qwen3-8.5B
Full fine-tune (fewer layers) vs Qwen/Qwen3-14B. Harvard-DCML/ADAPT-Qwen3-8.5B matches part of Qwen/Qwen3-14B (234 of its 443 tensors; the rest of Qwen/Qwen3-14B has no counterpart). On the shared part: Fine-tune: 100% of the weights changed, across 21 of 21 layers.
Tensors changed: Embeddings 1/1, Norms 85/85, MLP 63/63, Attention 84/84, lm_head 1/1
2 derivatives checked
- mradermacher/ADAPT-Qwen3-8.5B-i1-GGUF: Quantization
- mradermacher/ADAPT-Qwen3-8.5B-GGUF: Same weights, converted