JSONL format: rules and examples

The rules#

• 1UTF-8 encoding.
• 2One complete JSON value per line — almost always an object.
• 3Lines separated by \n. A trailing newline at the end of the file is recommended.
• 4No commas between lines and no outer [ ].
• 5Newlines inside text are escaped as \n, never written literally.

Example: support tickets#

• 1{"ticket_id": "00042", "channel": "email", "subject": "Refund request", "resolved": true}
• 2{"ticket_id": "00043", "channel": "chat", "subject": "Password reset", "resolved": false}

Example: chat fine-tuning records#

Many chat fine-tuning formats use one conversation per line with a messages array. Field names vary by provider, so confirm against their current documentation.

• 1{"messages": [{"role": "user", "content": "Is this in stock?"}, {"role": "assistant", "content": "Yes, two left."}]}

Common mistakes#

• 1Pretty-printing records across several lines.
• 2Single quotes instead of double quotes.
• 3Trailing commas after the last field.
• 4Wrapping the file in [ ] like a JSON array.
• 5Saving as UTF-16 or with stray characters from a spreadsheet export.

Keeping records consistent#

JSONL itself does not require every line to have the same fields. For datasets you usually want that, which is what a JSON Schema check is for.