PEFT/LoRA adapter for Qwen/Qwen2.5-3B-Instruct, fine-tuned on the slm-learning GAD-tool translation track.

CLI track at 3B base: same 783 vocab-anchored pairs that took v2 to 30/30 on Qwen2.5-1.5B. Open question: does scaling the base to 3B push HumanEval and GSM8K off zero (since the larger base has more latent capability) without hurting the saturated GAD-tools metric? Free experiment — we already have the data and v2's settings are known good. 3B base = ~6GB bf16 weights, easy fit on A100 40GB but blocks local inference at fp16 on 1660 Ti (would need int8). Decision pending: if this run lifts reasoning meaningfully, plan an int8 inference path.