Building TrueSET: training data for code repair where every example is proven by execution, not judged by another LLM — the broken code has to fail a hidden verifier, the fix has to pass it. I ship the eval so you can rerun the number instead of trusting it.

Solo and new at this. I'd rather be corrected than flattered.

TrueSET_Data@proton.me https://huggingface.co/datasets/TrueSET/verified-code-repair