Papers, model releases and downloads — controlled experiments with identical data and recipe; every number is reproducible and all evaluation data is public.
8 open-weight bases (7B–72B) × one 270,208-sample training set × one QLoRA recipe × one n=1,612 held-out: a monotonic capacity ladder, generation gains, teachability anti-correlated with base score, and a fully local 72B statistically tied with the web-connected frontier.
Read the paperQwen3-14B fine-tune: TCM 73.5% / general 80.0% (beats R1-Distill at 14B on both axes). Free for learning & research; commercial use requires authorization (Global South / Belt & Road preferential terms). License & SHA256 included.
Read the paperR1-Distill-Qwen-14B base: TCM 71.0% / general 73.7%, with syndrome-differentiation chain-of-thought preserved. Free for learning & research; commercial use requires authorization.
Read the paperR1-Distill-Qwen-32B base: TCM 76.3% / general 83.8%; review specificity 0.73 (the 32B capacity transition), reasoning chain preserved. Free for learning & research; commercial use requires authorization.
Read the paperFull-data fine-tune adapter on Qwen2.5-72B: TCM 85.3% / general 90.3%, statistically tied with the web-connected frontier. Requires your own 72B base to merge. Free for learning & research; commercial use requires authorization.
Read the paperDedicated safety-review model: sensitivity 0.48 / specificity 0.90 (knowledge models are unfit for review). Free for learning & research; commercial use requires authorization.
Read the paper