PolyData's picture
Upload README.md with huggingface_hub
b9b9658 verified
metadata
license: cc-by-4.0
task_categories:
  - tabular-classification
  - time-series-forecasting
tags:
  - polymarket
  - prediction-markets
  - blockchain
  - trading
  - orderbook
size_categories:
  - 1B<n<10B

Polymarket Trade Capture (through March 5, 2026)

Every filled order on Polymarket's CLOB (Central Limit Order Book) from launch through March 5, 2026. Sourced from on-chain events via the Goldsky indexer.

Overview

Metric Value
Date range 2022-11-21 to 2026-03-05
Partitions 1,194 days
Parquet files 1,643
Size on disk ~21 GB
Estimated total rows ~1-2 billion

Schema

Column Type Description
timestamp int64 Unix timestamp (seconds) of the fill
maker string Maker wallet address (0x...)
makerAssetId string Maker's asset token ID (USDC = "0", otherwise CTF token ID)
makerAmountFilled int64 Amount filled on maker side (6 decimal USDC or CTF shares)
taker string Taker wallet address (0x...)
takerAssetId string Taker's asset token ID
takerAmountFilled int64 Amount filled on taker side
transactionHash string On-chain transaction hash
date string Partition key (YYYY-MM-DD)

How to use

from datasets import load_dataset

# Stream without downloading everything
ds = load_dataset("PolyData/polymarket_trade_capture_5Mar2026", streaming=True)
for row in ds["train"]:
    print(row)
    break

# Load a specific date partition
import pyarrow.parquet as pq
ds = load_dataset(
    "PolyData/polymarket_trade_capture_5Mar2026",
    data_files="data/date=2026-01-15/*.parquet"
)

Interpreting the data

Buy vs. Sell: When makerAssetId == "0", the maker is paying USDC (buying tokens). When takerAssetId == "0", the taker is paying USDC.

Price calculation: price = usdc_amount / (usdc_amount + token_amount) where amounts are from opposite sides of the fill.

Token IDs: Each binary market has two token IDs (YES and NO). Map token IDs to markets using Polymarket's API: GET https://clob.polymarket.com/markets/{condition_id}

Amounts: Both USDC and CTF token amounts use 6 decimal places. Divide by 1,000,000 for human-readable values.

Data source

Raw OrderFilled events from the Polymarket CTF Exchange contract on Polygon, indexed by Goldsky. This is public blockchain data.

Partitioning

Data is Hive-partitioned by date (date=YYYY-MM-DD/). High-volume dates (typically >5M fills) are split across multiple parquet files (data_0.parquet, data_1.parquet).

Limitations

  • No market metadata (condition IDs, questions, categories). Join with Polymarket's API or CLOB endpoints.
  • No order-level data (only fills). Resting order placement/cancellation events are not included.
  • Token IDs are raw CTF ERC-1155 IDs. You need external mapping to determine YES vs NO side.
  • Coverage ends March 5, 2026. No updates planned.

License

CC-BY-4.0. This is public blockchain data. Attribution appreciated but not required.