Training a model

After you have generated a dataset and are happy with its topics and distribution, you can go ahead and start training a model on the dataset, either by clicking the "Train Model" button on the version overview page or by navigating to the Models tab and clicking "Create Model".

Start by adding a name & description for your model, then input a system prompt for the training process if your usecase requires it.

In most cases, you'll also need to provide a system prompt. This initial prompt is used in the training process to prime the model with its basic qualities, what it should/shouldn't do, etc.

Different base models have different strengths, weaknesses, and sizes. We try to offer a variety of base models that are relevant at any given time, so pick one that best fits your use case.

Finally, select the dataset versions that will be used to train the model. You can select as many datasets as you want, and optionally control the number of samples used from that dataset in the training process. If no sample size is provided, the entire dataset will be used.

Model training takes 30 minutes - 2 hours depending upon the size of your dataset and how long each of the prompts and responses are. If you have notifications enabled you will receive an email when the model is done training.