Climate-1000-Qwen-27b

Standalone BF16 model created by merging the best SFT adapter (checkpoint 230) into the original unquantized base weights at revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. The source adapter was trained with NF4 QLoRA; saved adapter tensors were FP32. Adapter source: verified final adapter.

Training completed 460 optimizer steps over two epochs. Best validation loss: 0.3242049217224121 at step 230; final step-460 validation loss: 0.32464906573295593. These losses were measured during quantized training, not remeasured on the merged model. The heldout benchmark remains provisional; this publication does not establish scientific accuracy. No inference evaluation has been performed on this merged export. Use the included tokenizer/chat template with enable_thinking=False, matching training. Model architecture: AutoModelForCausalLM; use compatible Transformers/vLLM versions.

Downloads last month
48
Safetensors
Model size
27B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Prience91/Climate-1000-Qwen-27b

Base model

Qwen/Qwen3.8-27B
Finetuned
(981)
this model
Quantizations
1 model