Fine-tuning in practice

Let’s now fine-tune an open-source model on our custom dataset. In this section, we will show an example that implements LoRA and QLoRA for efficiency. Depending on the hardware you have available, you can select the technique that best corresponds to your configuration.

There are many efficient open-weight models we can leverage for task or domain-specific use cases. To select the most relevant LLM, we need to consider three main parameters:

• License: Some model licenses only allow non-commercial work, which is a problem if we want to fine-tune for a company. Custom licenses are common in this field, and can target companies with a certain number of users, for example.
• Budget: Models with smaller parameter sizes (<10 B) are a lot cheaper to fine-tune and deploy for inference than larger models. This is due to the fact that they can be run on cheaper GPUs and process more tokens per second.
• Performance: Evaluating the base...