Tick-level quotes, trades and L2 depth from the 5-minute up/down markets on BTC, ETH, SOL and XRP. Parquet, one file per asset per UTC day, every manifest carrying a sha256.
Counted across the whole collection period, not just the days on sale. Refreshed once a day.
| Asset | 5-minute | 15-minute | Total |
|---|
Four caveats, up here rather than in a footnote. If any one of them matters to your use case, nothing else on this page will help.
Best bid, best ask, and that is the whole of it. There are level and size columns, but they exist to keep the schema stable across tables, and every row has them at 0. A reconstructed ladder is a different table — and it does not reach all the way back.
Fifteen days of the thirty-five have it. Everything before is quotes, trades and spot candles: still fine for flow and spread work, useless if you came to model the book. That gap is exactly why those days bill at 55% of a full one.
Up to 20 July the timestamp is ts — our receipt time, rounded to the second. Ordering inside a second is gone and cannot be recovered. If your model cares about sequence at that resolution, buy from 21 July on, where ts_ms is exchange time in milliseconds.
It runs from 95.8% to 100%, depending on the day. Each day carries its own figure rather than an average across the archive; the calendar shows it, and the price is computed from it.
Every purchase covers all four assets, and whatever streams existed on the dates you pick. Longer ranges are cheaper per day.
| Range | Rate | Example |
|---|---|---|
| 1–6 days | $5 / day | 3 days, $15 |
| 7–29 days | $4 / day | 10 days, $40 |
| 30+ days | $3 / day | 30 days, $90 |
Two rules keep the arithmetic honest. Days holding no data are not counted at all, so a range that straddles a gap costs less than its calendar length suggests. And a longer range never costs less than a shorter one — which is why the total flattens at $90 anywhere between 23 and 30 days.
Loading available dates.
The stripe under each date is that day's completeness: length is the share of streams present, amber means the collector lost time. You pay for the full-day equivalent, so a patchy stretch costs proportionally less. If the total surprises you, open the breakdown under the price — it shows every step of the sum.
One Parquet file per asset, per table, per day. Each ships with a manifest carrying its sha256, row count and interval bounds, so a download can be verified without taking our word for it.
| Table | Time column | Contents | Available |
|---|---|---|---|
| book_snaps | ts, ts_ms | Best bid and ask | All days |
| book_depth | ts_ms | Ladder levels: price, size, side. Around 88 per event. | From 20 Jul |
| clob_trades | ts | Trades with taker side | All days |
| clob_trades_ws | ts_ms | Trade stream with tx hash and fee, for VPIN | From 21 Jul |
| spot_klines | ts | Spot candles for the underlying | All days |
| book_meta | ts_ms | Market parameters. In practice only book_hash is populated. | From 21 Jul |
Four days — 26 to 29 June — are up for free. Read the schema and the file layout before you spend anything. Preview on Hugging Face
Payment goes through Plisio, in crypto. Once it confirms you get a download list good for thirty days, with a request budget sized to your order. The bytes come straight from storage and never pass through our server.
A month is several hundred files. Clicking through them by hand is nobody's idea of a good evening, so the order page hands over the whole list as one text file, with a ready-made command under it for Linux, macOS and PowerShell.
If the data disagrees with what this page says, you get your money back. That is the whole reason the caveats sit at the top and not down here in the small print.
The same stack that collects this dataset also trades on it. One payment, no subscription, no seats: source, an installer that takes a bare Ubuntu server to a running bot in about half an hour, and the pipeline that trains the models. The models themselves are not included — you train those on your own data, and they are yours.
Runs the whole loop against live market data without ever placing an order. What you are buying is the plumbing: already built, already running, not sketched.
Everything above, except it actually places the orders on the Polymarket CLOB.
Position size, how many trades at once, the daily loss cap, the entry price ceiling — all set from chat, applied to the running bot, no restart. The live edition takes your wallet the same way: a guided prompt that deletes the message with the key in it and stores the key encrypted.
artifacts/ arrives empty. Gate A will not let you train on less than two weeks of collected data, and until there is a model the bot runs, reports and holds. Buying a month of history from the calendar above is the only way to shorten that wait. A first Gate B verdict of FAIL is the ordinary outcome, not a defect.
Two cores and 4 GB of RAM are enough, but recording the full book is 25–40 GB a day. The docs cover keeping a window instead, or trimming what you record — and training is hours of CPU you should not run on the box that is collecting.
A tarball with the source, an idempotent install.sh, systemd units, the database schema in one file, and documentation written for someone who has not seen the code before. A manifest carries a sha256 for every file in it.
One buyer, any number of your own machines, source included and yours to modify. No redistribution and no running it as a service for other people. Support covers installation and defects — not strategy design.
Payment goes through Plisio, in crypto, same as the datasets. The archive is a single download from the same storage, and the link is good for thirty days — keep a copy of what you download.
Questions before buying? Message @Datasets_support_bot.