The rules are simple. The file is UTF-8 text. Every line parses as valid JSON without reference to any other line. A line break inside a text value is written as an escape sequence, since a real line break would end the record. Strict readers reject blank lines. Records should share the same keys and value types throughout.
The format suits training pipelines, which consume examples as a stream. A program can process a very large file without loading it into memory, and an error on one line does not make the rest unreadable. Nested content fits naturally: a conversation can be stored as a list of messages inside one object, each with a role and content. Typical faults are a JSON array saved under the wrong extension, pretty-printed objects that span several lines, unescaped quotation marks, trailing commas and inconsistent keys. All can be found before delivery with a validator. The free tools on this site convert CSV and JSON to JSONL and validate the result in the browser.
Compressed JSONL, usually with gzip, is common for large deliveries and keeps the line-by-line structure once unpacked.
Example
One line of a chat-format file holds a single conversation: an object with a messages list, in which each item has a role, such as user or assistant, and the text of that turn.