Large language models (LLMs for short) have taken the world by storm. They are machine learning models that are able to complete all sorts of tasks from content generation, to understand large documents, to writing code.

Mistral-7b has quickly become the leading model of choice for those working with open-source large language models (LLMs.) Released by Mistral AI, it outperforms the previous leading open-source models, even when they are much larger. It exceeds both Llama-2 13B and Llama-1 34B on all benchmarks, meaning Mistral-7B outperforms models that up to 6x larger!

You can fine-tune Mistral-7B on Forefront in about 30 seconds. It's as simple as uploading a dataset and clicking "Fine-tune". Forefront provides automatic loss charts, evals, and validation datasets out of the box. This lets both understand your fine-tuned model from both a quantitative and qualitative point of view. You can take your fine-tuned models and host them wherever you want. We look forward to seeing all the ways our customers will fine-tune Mistral-7B.

What is fine-tuning?

Fine-tuning is the act of teaching an LLM to perform well on your specific task. To do this, you must supply a dataset of examples that represent how you want your model to behave.

Example: Fine-tuning for résumé review

Let's say you want an LLM to review resumes for you. Specifically, you want the LLM to write the person's name, a summary of their background, a sentence on whether their background is aligned with the role, and a list of their largest achievements.

For this task you could create a dataset of examples where each example has an input (the prompt) and an output (the completion.) Your inputs would contain the text content of the resume, and the output would be a report for that resume.

Your dataset examples will likely come from past data you've collected, although it doesn't have to. Luckily, LLMs learn fast and you can likely get a good result with 50-100 examples for this particular use case.

How to fine-tune on Forefront

Next we'll see how to fine-tune a model on Forefront.

Prepare your dataset

Datasets are JSONL files that contain your examples. JSONL is a file format where each line is a JSON object. For fine-tuning each sample must be in either prompt-completion format or the conversation chat-ml format. For more information on formatting datasets, please refer to our documentation.

Start a new fine-tuning job

It's easy to fine-tune once you have your dataset:

Create an account and log in at https://forefront.ai
Click Fine-tune in the sidebar to go to the fine-tuning page
Click the Fine-tune button in the top right to create a new fine-tuned model

Add fine-tuning details

To create a fine-tuned model you'll need to provide at minimum four pieces of information:

a name for your model
the dataset to train on
the base model, or the model you want to fine-tune (in this case Mistral-7B)
the number of epochs to train for (how many times should the model train on all of the data in your dataset)

There are a few additional fields that you will probably find useful.

Validation datasets are useful as a qualitative way to view model performance. If you supply a validation dataset, we will inference your model on the examples in the dataset and you can view the results.

Evals are a great way to view how your model performs on common benchmarks. You can select one or more evals to run once your fine-tuned model has completed training.

Lastly, click "Fine-tune" and that's all! You'll be able to view training progress in the UI, and can try out the model in the playground or through API once it's ready.

Fine-tuning without Forefront

There are a number of alternative libraries that are used for fine-tuning such as transformers and axolotl. We recommend checking out their documentation and tutorials if interested.