Polymarket-v1

Four layers cover the full Polymarket v1 contract lifecycle from 2022-11-21 to 2026-04-28: OrderFilled/, the raw on-chain trade tape (13 columns); daily_aligned/, cleaned Standard Binary markets (24 columns); daily_aligned_multi/, cleaned Neg Risk multi-outcome markets (25 columns); and CTF/, the Conditional Tokens Framework lifecycle events — condition preparation, position splits and merges, resolution and payout redemption (5 to 8 columns per table). Each fill carries an execution price in USDC.e and a taker_direction field described as ground-truth direction from the settlement layer rather than a tick-rule or quote-rule inference, which is what makes the data usable for testing trade-classification methods instead of assuming one.

The card warns explicitly before any volume computation: OrderFilled/ is the nominal raw tape and can include platform relayer and router records, while daily_aligned/ and daily_aligned_multi/ are the cleaned layers with relayers removed — and the two cleaned layers must not be pooled without checking neg_risk. The cleaned layers are subsets of the tape, so the four configs’ row counts cannot be added together; the trade tape itself is 1,201,580,990 fills, a figure the README and the Hugging Face datasets-server state identically. That server reports 2,786,380,324 rows and ~52.7 GB of Parquet across all four configs. At this scale the card recommends streaming for row-level inspection and DuckDB or other columnar scans for aggregates, not full in-memory loads.

The CC-BY-4.0 license is declared in the Hugging Face frontmatter and restated in the README, but there is no LICENSE file in the repository — the only non-Parquet files are .gitattributes and README.md. The card also disclaims any affiliation with or endorsement by Polymarket and provides the data as is. A companion arXiv preprint, Polymarket-v1 Database (arXiv:2606.04217), has no peer-reviewed venue; its authors report 1.20 billion trade records across 1.30 million markets and $61 billion in nominal volume, and find that standard trade-classification methods achieve near-random accuracy in prediction markets. Its abstract gives a wider November 2021 start date than the dataset card, whose 2022-11-21 matches the earliest OrderFilled partition file.