PandaLM, an open-source "referee large model" from Peking University, West Lake University and others: three lines of code to fully automatically evaluate LLM, with an accuracy of 94% of ChatGPT

After the release of ChatGPT, the ecosystem in the field of natural language processing has completely changed. Many problems that could not be solved before can be solved using ChatGPT.

However, it also brings a problem: the performance of large models is too strong, and it is difficult to evaluate the differences of each model with the naked eye.

For example, if several versions of the model are trained with different base models and hyperparameters, the performance may be similar from the examples, and the performance gap between the two models cannot be fully quantified.

Currentlythere are two main options for evaluating large language models:

1. Call OpenAI’s API interface for evaluation.

ChatGPT can be used to evaluate the quality of the output of two models, but ChatGPT has been iteratively upgraded. The responses to the same question at different times may be different, and the evaluation results exist

Cannot reproduce problem.

2. Manual annotation

If you ask for manual annotation on the crowdsourcing platform,

the team with insufficient funds may Unable to afford it, there are also cases where third-party companiesleak data.

In order to solve such "large model evaluation problems", researchers from Peking University, Westlake University, North Carolina State University, Carnegie Mellon University, and MSRA collaborated to develop a PandaLM, a new language model evaluation framework, is committed to realizing a privacy-preserving, reliable, reproducible and cheap large model evaluation solution.

## Project link: https://github.com/WeOpenML/PandaLM

Provided with the same context, PandaLM can compare the response output of different LLMs and provide specific reasons.

To demonstrate the tool’s reliability and consistency, the researchers created a diverse human-annotated test dataset consisting of approximately 1,000 samples, in which PandaLM-7B’s accurate The rate reached 94% of

ChatGPT’s evaluation capability

. Three lines of code using PandaLM

When two different large models produce different responses to the same instructions and context, PandaLM is designed to compare the two large models. The response quality of the model, and output comparison results, comparison reasons, and responses for reference.

There are three comparison results: response 1 is better, response 2 is better, and response 1 and response 2 have similar quality.

When comparing the performance of multiple large models, you only need to use PandaLM to compare them in pairs, and then summarize the results of the pairwise comparisons to rank or draw the performance of multiple large models. The model partial order relationship diagram can clearly and intuitively analyze the performance differences between different models.

PandaLM only needs to be "locally deployed" and "does not require human participation", so PandaLM's evaluation can protect privacy and is quite cheap.

In order to provide better interpretability, PandaLM can also explain its selections in natural language and generate an additional set of reference responses.

In the project, researchers not only support using PandaLM using the Web UI for case analysis, but also for ease of use, It also supports three lines of code to call PandaLM for text evaluation generated by arbitrary models and data.

Considering that many existing models and frameworks are not open source or difficult to complete inference locally, PandaLM supports using specified model weights to generate text to be evaluated, or directly passing in a .json file containing the text to be evaluated.

Users can use PandaLM to evaluate user-defined models and input data by simply passing in a list containing the model name/HuggingFace model ID or .json file path. The following is a minimalist usage example:

In order to allow everyone to use PandaLM flexibly for free evaluation, researchers The model weights of PandaLM have also been published on the huggingface website. You can load the PandaLM-7B model through the following command:

Features of PandaLM

Reproducibility

Because the weights of PandaLM are public, even the output of the language model There is randomness. When the random seed is fixed, the evaluation results of PandaLM can always remain consistent.

The update of the model based on the online API is not transparent, its output may be very inconsistent at different times, and the old version of the model is no longer accessible, so the evaluation based on the online API is often not accessible. Reproducibility.

Automation, privacy protection and low overhead

Just deploy the PandaLM model locally and call the ready-made commands You can start to evaluate various large models without having to keep in constant communication with experts like hiring experts for annotation. There will also be no data leakage issues. At the same time, it does not involve any API fees or labor costs, making it very cheap.

Evaluation Level

To prove the reliability of PandaLM, the researchers hired three experts to conduct independent repeated annotations , a manually annotated test set was created.

The test set contains 50 different scenarios, and each scenario contains several tasks. This test set is diverse, reliable, and consistent with human preferences for text. Each sample of the test set consists of an instruction and context, and two responses generated by different large models, and the quality of the two responses is compared by humans.

Screen out samples with large differences between annotators to ensure that each annotator's IAA (Inter Annotator Agreement) on the final test set is close to 0.85. It is worth noting that the training set of PandaLM does not have any overlap with the manually annotated test set created.

These filtered samples require additional knowledge or difficult-to-obtain information to assist judgment, which makes it difficult for humans to Label them accurately.

The filtered test set contains 1000 samples, while the original unfiltered test set contains 2500 samples. The distribution of the test set is {0:105, 1:422, 2:472}, where 0 indicates that the two responses are of similar quality, 1 indicates that response 1 is better, and 2 indicates that response 2 is better. Taking the human test set as the benchmark, the performance comparison of PandaLM and gpt-3.5-turbo is as follows:

It can be seen that PandaLM-7B is in accuracy It has reached the level of 94% of gpt-3.5-turbo, and in terms of precision, recall, and F1 score, PandaLM-7B is almost the same as gpt-3.5-turbo.

Therefore, compared with gpt-3.5-turbo, it can be considered that PandaLM-7B already has considerable large model evaluation capabilities.

In addition to the accuracy, precision, recall, and F1 score on the test set, the results of comparisons between 5 large open source models of similar size are also provided.

First used the same training data to fine-tune the five models, and then used humans, gpt-3.5-turbo, and PandaLM to compare the five models separately.

The first tuple (72, 28, 11) in the first row of the table below indicates that there are 72 LLaMA-7B responses that are better than Bloom-7B, and there are 28 LLaMA The response of -7B is worse than that of Bloom-7B, with 11 responses of similar quality between the two models.

So in this example, humans think LLaMA-7B is better than Bloom-7B. The results in the following three tables show that humans, gpt-3.5-turbo and PandaLM-7B have completely consistent judgments on the relationship between the pros and cons of each model.

Summary

PandaLM provides a third article in addition to human evaluation and OpenAI API evaluation For solutions to evaluate large models, PandaLM not only has a high evaluation level, but also has reproducible evaluation results, automated evaluation processes, privacy protection and low overhead.

In the future, PandaLM will promote research on large models in academia and industry, so that more people can benefit from the development of large models.

The above is the detailed content of PandaLM, an open-source "referee large model" from Peking University, West Lake University and others: three lines of code to fully automatically evaluate LLM, with an accuracy of 94% of ChatGPT. For more information, please follow other related articles on the PHP Chinese website!