Data vendorslibraryfree
Polymarket_data (SII-WANGZJ): what it holds, what it costs, where it stops
Last checked 2026-09-08
Polymarket_data is a Hugging Face dataset from a team at the Shanghai Innovation Institute, Westlake University, Shanghai Jiao Tong University, Harbin Institute of Technology and Fudan University. The dataset card describes 163 GB of on-chain trading data, 1.9 billion records across 538K+ markets, fetched directly from Polygon. It has 21,501 downloads and 82 likes, was last modified July 21, 2026, and the companion GitHub toolkit has 825 stars.
It is for researchers who want the complete CLOB fill history rather than a vendor sample, and who can work with Parquet at this scale.
What it holds
Five files. trades.parquet (28 GB, 418.3M rows) is the recommended one: processed trades with market metadata linked. orderfilled.parquet (84 GB, 689.0M rows) holds the raw OrderFilled events. markets.parquet (85 MB, 538,587 rows) is metadata. quant.parquet (28 GB, 418.2M rows) restates trades from a unified YES perspective, and users.parquet (23 GB, 340.6M rows) splits activity by maker and taker role.
The card states coverage is every OrderFilled event from Polymarket's two CLOB exchange contracts with no missing blocks, from the first record on November 21, 2022 onward. The earlier FPMM/AMM era (2020 to November 2022) predates those contracts and is excluded.
- Fields the card contrasts with third-party sources: block_number, contract name, maker, taker and protocol fees, and order_hash.
- Download with hf download SII-WANGZJ/Polymarket_data using the dataset repo type, or one file at a time.
- The GitHub toolkit fetches new blocks continuously (batch when 100+ blocks behind, then one block every 2 seconds) and rebuilds the derived files.
- MIT license on both the dataset repo and the code.
Pricing
Free. Hugging Face hosts the files; the license is MIT.
Venues and data
Polymarket only, and only what settles on Polygon. Fills, fees and wallet roles are complete; resting orders, cancellations and depth never reach the chain and are absent. Each trade row links to market metadata, and the card states every trade was verified against blockchain RPC. Loading with pandas.read_parquet, or with the datasets library using a data_files filter, is documented on the card.
Where it falls short
No order book. No Kalshi. The GitHub pipeline was last pushed January 1, 2026 while the dataset was updated July 21, 2026, so the published code may lag the process that produced the newest files; the README on GitHub still quotes the older 1.1 billion record, 107 GB figures.
Polymarket moved to new CTF Exchange contracts on April 28, 2026 according to the poly_data maintainers. Confirm which contracts the July 2026 refresh indexed before treating post-April rows as complete.
Who should use it
Academics and quants doing behavioral or microstructure research on Polymarket who need every fill with fees and wallet roles.
Anyone who needs depth, Kalshi or minute bars without processing 163 GB should use DepthFeed, polymarketdata.co or the Becker dataset instead.
Questions about Polymarket_data (SII-WANGZJ)
Is SII-WANGZJ/Polymarket_data free?
Yes. It is MIT licensed and hosted on Hugging Face; the fetch and processing toolkit is on GitHub under the same license.
How far back does the Polymarket_data dataset go?
To the first CLOB OrderFilled event on November 21, 2022. AMM-era trades from 2020 to November 2022 are not included.
Does Polymarket_data include order book data?
No. It contains on-chain fills, market metadata and derived per-user and per-market views, not resting orders or depth.
How big is the Polymarket_data dataset?
163 GB across five Parquet files, 1.9 billion records; trades.parquet alone is 28 GB.
Official APIs for building. One book for trading.
Rather trade the spread than build the plumbing? The free Kairos terminal gives you Kalshi, Polymarket and Predict.fun in one order book.