candor lab

candor lab removes refusals from open models with targeted weight edits, and measures every release on the exact file you run.

Each model card reports how often the model still refuses, how far it drifts from the original, and whether its maths holds up. Everything is measured in llama.cpp, on the GGUF you download, using prompts that weren't used to make it.

The tool behind these models is called candor. Its code is open source under Apache 2.0: github.com/candor-lab-research/candor.

Requests and feedback are welcome: open a discussion in the Community tab. We read everything but can't promise every request.