Tracking LLMs with Comet

When building with LLMs, you will spend a lot of time optimizing prompts and diagnosing LLMs.

As you put your solutions into production, you need LLMOps tools to track LLMs and analyze prompts.

Here is a demo of how this process might look (use case included):

Step 1 - The Tool

I’ll use Comet new LLMOps functionalities to support our solution.

Comet ML provides prompt engineering tools to analyze prompts at scale.

Step 2 - Our Use Case

For the use case, I’ve built an ML paper tagger that uses ChatGPT-3.5 to extract the model names mentioned in paper abstracts.

After the LLM makes the tag predictions, we evaluate the accuracy of the results using another LLM (referred to as the LLM evaluator).

I’ve built a small validation dataset to evaluate the solution.

I'm experimenting with few-shot and zero-shot prompting. Here is an example of a zero-shot prompt to perform the tagging:

Step 3 - Creating LLM Project

I use Comet’s LLM tools to track how both the paper tagger and the LLM evaluator perform.

The first step is to create an LLM Project in Comet.

Step 4 - Logging Prompts in Comet

The next step is to use Comet’s LLM SDK to log all the prompts and related information to Comet. The functionalities are available via the open-source comet-llm Python library.

The comet-llm library helps to log prompts, responses, prompt templates, variables, and other related metadata you want to track.

For this use case, I am logging the prompts, the expected LLM response, the predicted LLM response, and the final verdict of the LLM-powered evaluator. Remember, we are interested in assessing and tracking the performance and quality of the LLM evaluator so we also want to create helpful tags to help us track this easily.

Here is an example of the prompt information and metadata I am logging:

Step 5 - Analyzing LLM Behavior

We are particularly interested in analyzing the LLM evaluator so it’s useful to filter by the final verdict tag (INCORRECT/CORRECT) and other bits of information to take a closer look at the behavior and quality of the LLM-powered evaluator.

Comet 's LLMOps functionalities allow us to quickly track and get insights about the effectiveness of the LLM evaluator.

We can easily navigate prompts/responses and metadata. We can also filter and group the prompt logs to get faster overall trends/insights.

Having the prompt logs in Comet allows us to track the performance of the LLM evaluator in real-time which is super useful.

Step 6 - Search for Prompts

You can also perform quick searches on your prompt logs via the interface.

For our use case, we are interested in quickly finding specific keywords, errors, or prompts of interest.

I search for relatively new ML concepts (Llama 2 or GPT-4) as I suspect the model might struggle to extract these names from the paper abstracts.

From a few searches and navigating the logs, I have managed to gather a few insights and have a better understanding of the behavior of the LLMs.

For instance, I observed that the LLM evaluator is misclassifying the prompt responses as INCORRECT even when the predicted tags overlap with the expected tags in the evaluation dataset. Here is an example of this behavior:

Next Steps

There is more work to do on improving the LLM evaluator itself. I can continue iterating and improving my solution and use Comet to continue monitoring the quality of the LLM evaluator.

The next step would be to potentially experiment with better LLM evaluators or directly improve the prompt template used by the LLM evaluator chain.

This is a simple use case, but you can use Comet LLMOps functionalities to track and debug your LLMs for a wide range of use cases. The team is working on many other features such as tracking user feedback, better grouping features, and viewing and diffing prompt chains.

If there is enough interest, I might do a follow-up thread to demonstrate other steps we can take to keep improving our solution and share more insights.

Find the notebook I used for this demo here: https://github.com/dair-ai/llm-evaluator

Comet LLM SDK: https://github.com/comet-ml/comet-llm

Docs: https://www.comet.com/docs/v2/guides/large-language-models/overview/