Projects

Two repositories, both Apache-2.0, both maintained in the open.

conform Live

A test suite that sends identical requests to vLLM, llama.cpp, Ollama and SGLang and reports where their answers differ. It found that Ollama ignores max_tokens=0 and generates an unbounded response.

ladder In progress

A QLoRA fine-tuning pipeline for Codeforces problems that runs on a free Kaggle T4. Generated programs are executed against the problem’s own test cases, so the score is pass rate rather than a model’s opinion.