This is the code repo for EMNLP 2024 paper "Bayesian Calibration of Win Rate Estimation with LLM Evaluators".
Recent advances in large language models (LLMs) show the potential of using LLMs as evaluators for assessing the quality of text generations from LLMs. However, applying LLM evaluators naively to compare different systems can lead to unreliable results due to inaccuracy and intrinsic bias of LLM evaluators. In order to mitigate this problem, we propose two calibration methods, Bayesian Win-Rate Sampling (BWRS) and Bayesian Dawid-Skene, both of which leverage Bayesian inference to more accurately infer the true win rate of generative language models. We empirically validate our methods on six datasets covering story generation, summarization, and instruction following tasks. We show that both our methods are effective in improving the accuracy of win rate estimation using LLMs as evaluators, offering a promising direction for reliable automatic text quality evaluation.
Our datasets can be accessed in our huggingface organization. These datasets will be accessed by modules under dataset_loaders, there's no need to download them manually.
First, install the required packages:
pip install -r requirements.txt
All the experiments can be replicated using the scripts under scripts/. To run an experiment, please use:
bash scripts/<method name>/<prior setting>/<dataset name>.sh
where method name can be one of bayesian_ds or bwrs, prior setting can be one of in_dist, no_prior, or ood_prior for Bayesian Dawid-Skene, and one of in_dist or ood_prior for BWRS. For example, to run the Bayesian DS experiment with no prior, please use the following command:
bash scripts/bayesian_ds_no_prior.sh
The script will run the experiment and save the results to ./results/. To process the results, please refer to the "process logs" section in experiments.ipynb. Specifically, plot_all_in_dist_error_curves can be used to produce curves similar to Figure 2, and parse_repeat_results can be used to generate result summaries similar to Table 2 and 3 in our paper.
This project is licensed under the Apache License 2.0 - see the LICENSE file for details.
If you use our code, data, or methodology in your work, please cite our paper: