A better method for identifying overconfident large language models

Build a standalone Python library for measuring LLM calibration via 'verbalized confidence' and logit analysis. It should integrate seamlessly with HuggingFace models to provide an 'Uncertainty Score' for any generation.

Suggested repo: caliLLM

"Stop trusting hallucinating models—measure their certainty."

Estimated effort: 60h