Your data is read and written on this device. Nothing is uploaded.
Chat and completion formats
Chat datasets have one JSON object per line with a messages list: an optional system message, then user and assistant turns, ending with the answer you want the model to learn. Completion datasets have a prompt and a completion on each line. Most current fine-tuning services expect the chat format; some evaluation tools and older services use prompt and completion.
What the checker looks for
Lines that aren't valid JSON, lines that aren't one object, missing or empty messages, unknown roles, empty content, examples with no assistant answer, system messages in the middle, the same role twice in a row, tool messages without a tool_call_id, mixed formats, duplicate lines and blank lines. Errors usually make an upload fail; warnings are worth a look but may be fine.
Questions
How should my CSV look?
One example per row, with a column for the prompt and one for the answer, and optionally a column for a system message. Columns named prompt, question or input and completion, answer or output are picked up automatically; you can change the choice from the lists.
How many examples do I need?
Providers usually ask for at least ten, and results tend to improve with fifty to a few hundred clear, consistent examples. Quality matters more than quantity.
Are the token counts exact?
No, they're estimates at about four characters per token plus a few tokens for each message, which is close enough to judge cost and spot examples that are far too long.
Can it check a very large file?
Yes. Files with tens of thousands of lines check in a moment. The report lists the first 1,000 problems on screen and every problem in the CSV.