Dataset Splitter

Paste a JSONL fine-tuning file or any line-based dataset, choose your train/val/test percentages, shuffle reproducibly, optionally stratify by a label key, and download the three splits. Everything stays in your browser.

Your dataset (one example per line - JSONL or plain text)

Shuffle Seed Stratify by key Key

Train 80%

Val 10%

Test 10%

Total lines

Train

Val

Test

Why stratify? A plain random split can accidentally put most of a rare class into one split - for example, all your "refund" examples land in test and the model never trains on them. Stratifying by a key (like label or category) splits each value's examples separately with the same percentages, so every split keeps roughly the same class distribution as the full dataset. The seed makes the shuffle reproducible: the same data, seed, and percentages always produce the same split, which matters when you re-run experiments. Files download as .jsonl when every line parses as JSON, otherwise .txt.