LLM Judge: Detecting hallucinations in language models

This project demonstrates how to fine-tune and evaluate a Mistral AI language model to detect factual inconsistencies and hallucinations in text summaries. It is based on this amazing blog post by Eugene Yan.

In this project, we will:

• Prepares datasets from Factual Inconsistency Benchmark (FIB) and USB
• Fine-tunes a Mistral 7B model for hallucination detection
• Evaluates model performance using accuracy, F1 score, precision, and recall
• Integrates with Weights & Biases for experiment tracking

Weave

In this project we make extensive use of Weave to trace and organize our model evaluations.

• You can get started with Weave and MistralAI by following the quickstart guide

Usage

1. Prepare the data:

   • Run 01_prepare_data.ipynb to process and format the datasets

   The dataset is also available in the data folder, so you may skip this notebook.

2. Fine-tune and evaluate the model:

   • Run 02_finetune_and_eval.ipynb to:
     • Evaluate baseline Mistral models (7B and Large)
     • Fine-tune a Mistral 7B model
     • Evaluate the fine-tuned model

Results

The notebook demonstrates improvements in hallucination detection after fine-tuning, with detailed metrics and comparisons between model versions.

All the results and evaluation are logged to this Weave Project

The finetuning process is logged to Weights & Biases as well, living together on the same project as the model evals.

Docs

• Weights & Biases: https://wandb.ai/
• Mistral finetuning docs: https://docs.mistral.ai/capabilities/finetuning/
• Tracing with W&B Weave: https://wandb.me/weave

Notes

• Ensure you have the necessary API keys for Mistral AI and Weights & Biases
• Adjust NUM_SAMPLES in the evaluation notebook to control the number of examples used

For more details, refer to the individual notebooks and comments within the code.