dphn/dolphin-2.9.1-llama-3-70b
LoRA-style vs meta-llama/Meta-Llama-3-70B. Looks like a merged LoRA (or a fine-tune with frozen norms): only the weight matrices of down_proj, gate_proj, k_proj, o_proj, q_proj, up_proj, v_proj changed, in 77 of 80 layers; norms, biases and embeddings are untouched.
Tensors changed: Norms 0/161, MLP 108/240, Attention 144/320
9 derivatives checked
- mradermacher/dolphin-2.9.1-llama-3-70b-i1-GGUF: Quantization
- legraphista/dolphin-2.9.1-llama-3-70b-IMat-GGUF: Same weights, converted
- mradermacher/dolphin-2.9.1-llama-3-70b-GGUF: Quantization
- Nexesenex/Llama_3.x_70b_Tess_Dolphin_128K_v1.2: Full fine-tune
- Nexesenex/Llama_3.x_70b_L3.3_Dolphin_128K_v1.02: Full fine-tune
- NexesMess/Llama_3.x_70b_Tess_Dolphin_128K_v1.01: Full fine-tune
- DarqueDante/LLama-3-Mega-Dolphin-2.9.1-120b: Partial fine-tune (fewer layers)
- Nexesenex/Llama_3.x_70b_FLDx2_Dolphin_128K_v1.02: Full fine-tune
- NexesMess/Llama_3.x_70b_Flipper_0.34: Full fine-tune
3 other repos hold byte-for-byte identical weight files.