Hallucination Risk Calculator
a day ago
- #calibration
- #hallucination-risk
- #LLM
- Post-hoc calibration toolkit for large language models without retraining.
- Provides bounded hallucination risk using Expectation-level Decompression Law (EDFL).
- Supports two deployment modes: Evidence-based and Closed-book.
- Uses OpenAI Chat Completions API for scoring without model retraining.
- Includes mathematical framework for information budget and prior masses.
- Decision to ANSWER or REFUSE based on Information Sufficiency Ratio (ISR).
- Offers practical considerations, calibration adjustments, and deployment options.
- Comes with installation instructions and project layout details.
- Includes API surface details and validation methods.
- Licensed under MIT License by Hassana Labs.