Free · no signup · browser-local
JSONL Train, Validation and Test Splitter with Seed
Split a JSONL dataset into reproducible train, validation and test files. Set ratios, a random seed, duplicate handling and a group field, then download the ZIP.
Browser-local processing · no model calls · no input storage
Your output
train.jsonlProcess the input to preview, copy and download a result.
Local preparation only. No remote fetching, code execution, model grading or platform eligibility checks. Editing an input invalidates previous results and pending file exports.
Getting started
Load JSONL, set three percentages adding to 100 and enter a seed. Optionally choose an exact top-level group field, then inspect actual counts and choose seeded shuffle or source order, optionally preserve source order inside each partition, and download three files with a content-free summary manifest ZIP.
Before you use the result
Groups never cross partitions, so actual ratios can differ from requested ratios. This does not detect semantic duplicates or guarantee a statistically representative split. Duplicate removal compares exact raw lines and is optional.
Try an example
Ten ungrouped records at 60/20/20 produce 6 training, 2 validation and 2 test records. Reuse the same input, seed and options to reproduce the result.
How it works
Seeded ordering and explicit allocation with atomic groups. Original JSON line text and numeric spellings are preserved, with LF file separators. Inputs are limited to 500 KB UTF-8 and 1,000 records; system instructions to 8 KB and generated JSONL to 5 MB.
Frequently asked questions
Split JSONL reproducibly with a seed, explicit ratios, optional duplicate removal and a field that keeps groups together.
How do I use AI Dataset Train / Validation / Test Splitter?
Load JSONL, set three percentages adding to 100 and enter a seed. Optionally choose an exact top-level group field, then inspect actual counts and choose seeded shuffle or source order, optionally preserve source order inside each partition, and download three files with a content-free summary manifest ZIP.
Does this tool upload my data?
Inputs and files are processed in your browser. Page-view analytics do not contain your tool input.
What should I check before using the result?
Groups never cross partitions, so actual ratios can differ from requested ratios. This does not detect semantic duplicates or guarantee a statistically representative split. Duplicate removal compares exact raw lines and is optional.