--- base_model: Qwen/Qwen3.5-9B library_name: peft license: cc-by-sa-4.0 tags: [lora, decision-model, openchoice] --- # OpenChoice-9B decision adapter Rank-16 LoRA (alpha 32) for the native slot decision reader, based on `Qwen/Qwen3.5-9B` revision `c202236235762e1c871ad0ccb60c8ee5ba337b9a`. The selected adapter was trained for 100 updates on 3,200 rows / 3,234,666 tokens, seed 20260921. This directory preserves the original adapter weights for fine-tuning and reproducible merging. It is not the serving package by itself. Use the GGUF files, processor, and manifest in the parent directory for CPU inference; use the public source's FINETUNING.md to continue training and export a new GGUF package. The slot reader produces Choice/Score/Noul probabilities from final-position answer logits without text generation. The parent `heads.safetensors` is empty because this checkpoint has no separate trained classifier head. See the parent model card, LICENSE, LICENSE-QWEN, and NOTICE for evaluation, limitations, attribution, and the publisher-owned data-rights review. The conservative CC-BY-SA-4.0 weights declaration is not a legal determination that source-data terms have been resolved. Code and Qwen base licensing are separate.