Browser-local AI data utility
JSONL Dataset Validator
Locate broken JSON Lines records, check a common dataset shape, and estimate token volume with an assumption you can see and change.
Browser-local AI data check
Validate JSONL before training or import
Choose a file or paste JSONL to begin.
Line-level inspection
Issues to fix
Dataset content stays on this device. Parsing and export run in the browser.
Planning estimate, not a tokenizer. Confirm exact tokens with the target model or provider before a paid run.
What validation proves
Every non-blank line must parse independently as a JSON object. The selected shape check then verifies required top-level fields and basic value types. Errors retain physical line numbers so you can return to the source file.
FAQ
Is my JSONL dataset uploaded?
No. The file is read and checked in your browser. Record content, filenames, errors, and token estimates are not sent to Useful Atlas analytics.
What shapes can the validator check?
It supports generic JSON objects, chat records with a messages array, prompt and completion pairs, and input and output pairs. These are structural checks, not provider-specific guarantees.
Is the token estimate exact?
No. It divides character count by your editable characters-per-token assumption. Exact counts depend on the tokenizer, model, language, whitespace, and content.
What is included in the valid-lines download?
Only non-blank lines that contain valid JSON and pass the selected structural check. The original line text is retained.