Climate-1000-Qwen-27b-Adapter API & Inference Endpoint

Climate-1000-Qwen-27b-adapter

Final best-validation adapter from the SFT run started 14 October 2026. Training completed all 460 optimizer steps / two epochs. The selected weights are checkpoint 230, with validation loss 0.3242049217224121. The final adapter's tensor values were checked against that checkpoint.

Base: Qwen/Qwen3.8-27B, revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. Dataset: Prience91/Climate-1000-canary, revision 18d438de908614cbe34ffcd91a563b3e8e320cc9. 3,665 train and 406 validation examples; all 100 reserved held-out paper identities excluded from SFT. The 78 approved test references remain provisional. Selection uses validation loss; no independent benchmark or deployment-quality claim is made.

QLoRA: frozen NF4 4-bit base with double quantization and BF16 compute; LoRA rank 32, alpha 64, dropout 0.05, all supported linear projections. Saved adapter tensor dtypes: F32. Compute precision and saved adapter precision need not be the same. Completion-only loss, thinking disabled, no packing, max length 8,192; microbatch 1, accumulation 16, learning rate 1e-4, seed 3407.

This repository contains the adapter, tokenizer, chat template, and checksum manifest. Load the pinned base in a compatible Transformers/PEFT environment and attach this adapter. For the training base representation use NF4, double quantization, and BF16 compute. Apply the chat template with enable_thinking=False. The adapter config pins the public base revision.

This is not a merged model. No merge, inference launch, or new evaluation is performed by publication. Optimizer/RNG state and serialized training arguments are excluded; full training artifacts remain on Spark. The previous intermediate upload remains untouched in the separate repository Prience91/GIIS-Mini-Qwen3.8-27B-QLoRA-adapter, tag checkpoint-230.

Climate-1000-Qwen-27b-Adapter

README