Train, validation and test splitter
Load a CSV, TSV, JSON or JSONL file and split it into train, validation and test sets by percentage. A seed makes the split repeatable. You can stratify by a class column, keep every row that shares a key in one split, or split by date. Everything runs in your browser.
How to use it
- 01Load a file, or try the sample of support tickets.
- 02Set the three percentages and a seed. The same file, settings and seed always give the same split.
- 03Pick a group key such as customer_id or thread_id so related rows stay together. Pick a class column to stratify, or a date column for a time-based split.
- 04Check the size of each split and the class balance table.
- 05Download each file, or all three as a zip, in the format you loaded.
What it does not do
- With a group key, split sizes are approximate because whole groups move together. One very large group can skew them.
- A time-based split ignores shuffle and stratify. Rows with a date that cannot be read are placed with the oldest rows.
- Stratifying a class with only a few rows cannot give every split a copy. Check the balance table.
- Files are read into memory, so the limit is 50 MB.
Questions
About the train, validation and test splitter.
Is my data uploaded anywhere?
What is leakage, and how does a group key prevent it?
When should I split by time?
Will I get the same split next time?
Is it really free?
More free tools
Personal data scrubber
Strip names, emails, phones, cards, keys and addresses from text and files.
Dataset profiler
Rows, columns, tokens, date span, duplicates and personal-data columns in one report.
Column hasher and remover
Drop sensitive columns or replace them with salted hashes that still join.
CSV, JSON and JSONL converter
Convert between CSV, TSV, JSON and JSONL without breaking quoted cells.
JSONL validator
Find the broken lines in a JSONL file and see which keys are missing.
Duplicate remover
Remove repeated rows, records or log lines, keeping the first of each.
Token counter
Size text the way AI labs do: tokens, words, characters and lines.
Data card generator
Write the one-page description every data buyer asks for.
Data rights pre-check
A tailored checklist of the contracts and questions to clear before a sale.
Slack export parser
Turn a Slack workspace export into clean thread JSONL with pseudonymized users.
Email thread parser
Convert .mbox and .eml files into threaded JSONL without quoted replies.
Ticket export parser
Map help desk exports to customer and agent turns in conversation JSONL.
Chat dataset format converter
Convert and validate OpenAI, ShareGPT, Alpaca and prompt/completion datasets.
Log parser
Turn web server, syslog, logfmt and JSON logs into structured JSONL or CSV.
HTML and document text extractor
Extract clean text from HTML, Markdown, XML and JSON files in batches.
Re-identification risk checker
Measure k-anonymity and l-diversity, then preview fixes that lower the risk.
Secrets scanner
Find API keys, tokens, passwords and private keys in a repo or export before you share it.
Delivery manifest builder
Create a SHA-256 checksum manifest for a dataset delivery, with one hash to compare.
Delivery verifier
Check received files against a manifest: match, changed, missing or extra.
File checksum calculator
SHA-256 of any file, with a compare box that turns green or red.
Dataset sampler
Pull a random, stratified or first-N sample from a file, including files too large to open.
Schema inferrer and data dictionary
Infer field types and stats from a file and export JSON Schema, a data dictionary and SQL.
Text chunker
Split documents into overlapping chunks for retrieval or training, with heading context.
Dataset diff
Compare two versions of a dataset: added, removed and changed records, plus schema changes.
Encoding and mojibake fixer
Repair garbled characters, convert files to UTF-8 and strip invisible characters.
Column toolkit
Rename, split, merge, clean, filter and sort columns with a reusable pipeline of steps.
Cleaned it up? Sell it to Outrich.
Outrich Data buys company data directly. Check in five minutes whether yours qualifies, with no call.