Models

Models are what we came here to work with in the first place! And now, you should have enough background knowledge to customize one.

But first, let's define a few types of models, starting with what a "base model" is.

Base Models

Base models are what come off-the-shelf through our model providers. There are two types of base models:

Pre-trained Models

Pre-trained models have learned how to predict text, but not how to follow instructions. They will naturally try to auto-complete your prompt and are unpredictable in nature. Not very many providers release models in this state anymore.

Instruction-tuned Models

These models have been fine-tuned to follow instructions, which is pretty amazing when you think about it. The pre-trained model is fed examples of prompt/completion pairs (actually, system/prompt/completion trios) and learns to use its foundational understanding of language to follow instructions. We can steer it with words!

Templated Models

Templated models let you run a prompt (in the form of a template) on the base model of your choice. This is simply a way to do inference with a particular template and model (and evaluate it).

In Entry Point AI, creating a templated model is called "applying a template."

Fine-tuned Models

Fine-tuning a model starts a process that can take some time, and usually incurs cost from the model provider. Not all base models are available for fine-tuning.

When you start a fine-tune, Entry Point packages up your training examples (typically into a JSONL) with your template and sends them to the model provider via their API, which makes a copy of the base model you selected, and then starts to train it in batches on your examples. In the end, you have a customized version of the original model, trained specifically for your use case.

Fine-tuning is a very powerful technique, and Entry Point is unique because you can compare instructional prompts directly against fine-tunes, using both templated and fine-tuned models.