ZihanLiummyycc/FSG-RL-Stage2-SFT

FSG-RL executable SFT adapter

This is the Stage-2 executable-solving LoRA adapter for Qwen/Qwen3.5-9B-Base. It follows graph-construction SFT and trains the policy to produce tagged Python implementations for public function-graph nodes. The Stage-1 checkpoint is not part of this release; this Stage-2 checkpoint is the saved adapter used for evaluation.

It is adapter-only: the Qwen base weights must be obtained separately. Load it as a PEFT adapter on that base using compatible Transformers and PEFT versions. No optimizer state is needed for inference. The code, data schema, and historical configuration are in FSG-RL.

On the paper's 400-item custom graph-conditioned evaluation set, this checkpoint achieved 43.25% final-answer accuracy and 32.25% strict full success. Local evaluation used greedy decoding, thinking disabled, and a 2,048-new-token limit. These are not official scores on the source datasets; the public dataset export excludes private scoring fields.