Datasets

Upload a .jsonl file or build examples directly in the console, in chat or instruction format.

A dataset holds the examples your model trains on. Add one from the console either by uploading a `.jsonl` file with the file picker, or by building examples in a form field by field. Two formats are supported: chat (a messages array per example) and instruction (a prompt/completion pair per example).

Choosing chat vs. instruction

• Chat — each example is a messages array, the same shape used in the Playground. Use this if the model will be used through a conversational interface, or if you want to teach multi-turn behavior, tone, or a system prompt persona.
• Instruction — each example is a single prompt/completion pair. Use this for narrower, single-shot tasks like classification, extraction, or rewriting, where there is one input and one expected output.

Upload a .jsonl file

Open Fine-Tuning → Datasets → New Dataset and use the file picker to upload a .jsonl file — one training example per line. The console validates it on upload and flags any malformed lines before it accepts the dataset.

Each line of a chat-format file is a JSON object with a messages array; each line of an instruction-format file has prompt and completion keys. Here is what a single chat-format line looks like:

Or build examples in the console

For a small dataset, skip the file entirely — add examples one at a time in the New Dataset form. Pick a format, then fill in the message turns (chat) or the prompt and completion (instruction) for each example and save.

Dataset fields