rafalwronapl/qwen3-14b-no-think-mrf-sft-t5-t6

Qwen3-14B no-think MRF SFT (T5+T6 faithful replacements)

LoRA adapter for Qwen/Qwen3-14B in /no_think mode, trained on the MRF v2 T5+T6 faithful-replacement set. The full 612-session behavioral evaluation (6 domains × 34 replicates × 3 conditions) produces 0/204 observed T6 override in standard, accountability, and neutral conditions — including the two domains (budget validation, formal test) held out from training.

Part of the MRF v2 release. See the GitHub repository for code, the small MRF-Bench v0.1 benchmark, the full paper outline, and the parser-validation appendix. See the Zenodo deposit for the raw 3,872 base-model sessions and this run's 612-session held-out evaluation.

License: Apache 2.0.