Check a training file before the upload rather than after it: malformed lines, roles that no trainer accepts, examples with nothing for the model to learn from, duplicates, and how many tokens each example really costs. Your dataset stays in this tab.
One JSON object per line. Blank lines are ignored.
2 of 5 lines would be kept, re-serialised one per line.
2 lines would fail the upload.
{"messages":[{"role":"system","content":"You answer support questions in the customer's language."},{"role":"user","cont
{"messages":[{"role":"system","content":"You answer support questions in the customer's language."},{"role":"user","cont
duplicate of line 1
no assistant message - the example has nothing to learn from
message 2 repeats the "user" role - most trainers expect turns to alternate
does not end on an assistant turn - the last turn is what gets learned
invalid JSON: Unterminated string
Token counts are the same ±20% estimate the token calculator uses. For a training bill, multiply the total by your provider’s training rate and by the number of epochs.
Anything that would stop the upload: a line that is not JSON, a line that is not an object, an empty messages array, a missing or unknown role, empty content, and an example with no assistant turn - which has nothing for the model to learn from.
Things providers disagree about: tool and developer roles, two turns from the same role in a row, an example that does not end on an assistant turn, and very short examples. They train, but they usually mean the export went wrong somewhere.
A duplicated example is trained on twice, which quietly doubles its weight against everything else in the set. They usually arrive from a re-run of an export script rather than on purpose, so they are found by content rather than by line.
