kddubey/cappr

Computes conditional log-probabilities of completions given prompts, enabling classification tasks by ranking candidate answers. Supports both local models (via llama-cpp and Hugging Face transformers) and integrates prompt-prefix caching to optimize repeated computations across similar prompts. Provides batch processing, prior incorporation, and token-level probability aggregation for applications like DPO training and chain-of-thought answer extraction.

Used by 1 other package. No commits in the last 6 months. Available on PyPI.