LoRA adapters from three fine-tuning rounds attempting to improve Qwen2.5-Coder-14B-Instruct (4-bit MLX) at OpenQASM circuit generation, trained and evaluated with qcbench / qicode on the verified-openqasm-circuits dataset.

Headline: all three rounds are benchmark-neutral. Verified-rate on the qcbench extended suite is unchanged in every configuration (identical verified problem sets); validation loss converges to ~0.02-0.06 while circuit correctness does not move — the adapters learn the corpus style, not circuit semantics. We publish them because negative results with a clean instrument are worth more than silent failure.

Use with MLX: mlx_lm.server --model mlx-community/Qwen2.5-Coder-14B-Instruct-4bit --adapter-path round3/

Open question these adapters motivate: does FP16 higher-capacity SFT or verified-reward RL (GRPO on qcbench's dense scores) break this ceiling? That experiment needs GPU compute.