Fine-Tuning

Fine-tuning is incredibly powerful. PromptLayer lets you build and iterate models in a few clicks.

If you are already logging your gpt-4 requests in PromptLayer, it only takes a few clicks to fine-tune a gpt-3.5-turbo model on those requests! ✨

​

What is fine-tuning?

Fine-tuning is a technique to specialize a pre-trained large language model (LLM) for a specific task. It involves training the LLM on a small dataset of examples, where the input is the text to be processed and the output is the desired output, such as a classification label, a translation, or a generated text.

Fine-tuning is powerful because it allows developers to create a model that is tailored to their specific needs. This could be used to improve model output quality, shorten a system prompt without degrading performance, or to decrease latency by building off of a smaller model.

Here are some examples of how fine-tuning can be used:

• Reduce latency and cost: Fine-tune gpt-3.5-turbo on gpt-4 outputs to achieve gpt-4-quality results on a faster and cheaper model.
• Save on tokens: Generate training data using a long and complex prompt. When fine-tuning, change the prompt to something shorter and save on tokens.
• Improve output format: Generate synthetic training data to teach a base model to only output text in JSON.

​

Create training data

The first step to fine-tuning is preparing the training data you want the model to learn from. Training data in this case are just LLM requests.

​

Log in the background

The simplest way to do this is to just connect your application to PromptLayer and start logging requests. Just wait a week and your production users will have created tons of training data for you!

​

Batch run prompts

Alternatively, you can use PromptLayer to generate these training requests. Visit the Evaluations page to run batch jobs of your prompts.

For example, to generate fine-tuning data you can run a prompt template from the Prompt Registry against 200 test cases on gpt-4. Then just filter the sidebar based its specific test run tag.

​

Select training data

Use the sidebar search area to filter for your training data. All the data that appears from that search query will be used to fine-tune.

Learn more about search filters

​

Start the fine-tune job

Click “Fine-Tune” in the sidebar, follow the steps, and kick off a job.

​

Test out your new model

Success! 🎉 Now you have a new fine-tuned model. Let’s see if it’s any good…

​

Try it in Playground

Copy the model name and navigate to the PromptLayer Playground. There you can run an arbitrary request on the new model. See how it does!

​

Try it in Evaluations

It’s important to test your fine-tune model a little more rigorously than one-off Playground requests. Navigate to the Evaluations page and run some batch tests. See how the fine-tuned candidate compares to a standard gpt-4 candidate.