Add dataset card

README.md

@@ -0,0 +1,234 @@
+---
+license: cc-by-4.0
+language:
+  - en
+pretty_name: Polymarket Geopolitics — Markets, Trades, Blocks
+size_categories:
+  - 10M<n<100M
+task_categories:
+  - other
+tags:
+  - polymarket
+  - prediction-markets
+  - geopolitics
+  - blockchain
+  - polygon
+  - market-microstructure
+configs:
+  - config_name: markets
+    data_files:
+      - split: train
+        path: "polymarket/markets/markets_*.parquet"
+  - config_name: trades
+    data_files:
+      - split: train
+        path: "polymarket/trades/trades_*.parquet"
+  - config_name: blocks
+    data_files:
+      - split: train
+        path: "polymarket/blocks/blocks_*.parquet"
+---
+
+# Polymarket Geopolitics — markets, trades, blocks
+
+A curated subset of [Polymarket](https://polymarket.com) markets and on-chain
+trade events covering the **Geopolitics** category. Polymarket is a
+decentralised prediction market on the Polygon blockchain; binary
+contracts settle to one USDC if the underlying event occurs and zero
+otherwise. The dataset is intended for teaching and research on price
+discovery, calibration, and trader behaviour in prediction markets.
+
+## Contents
+
+| File group | Rows | Description |
+|---|---:|---|
+| `polymarket/markets/markets_*.parquet` | 6,053 | One row per Polymarket question classified as Geopolitics. |
+| `polymarket/trades/trades_*.parquet` | 11,734,433 | One row per CTF-Exchange `OrderFilled` event for tokens belonging to those markets. |
+| `polymarket/blocks/blocks_*.parquet` | ~12 M | Polygon `block_number` → `timestamp` lookup, sampled every 100 blocks and linearly interpolated in between. |
+
+Each file group is chunked into 10,000-row parquet shards (1,174 trade
+shards, 785 block shards, 1 market shard) — same chunking convention as
+the upstream release.
+
+## Provenance and attribution
+
+The raw market metadata, trade events, and block index were collected and
+released by **Jon Becker** at
+[github.com/jbecker19/polymarket-data](https://github.com/jbecker19/polymarket-data).
+This dataset is a category-restricted derivative of that release: all
+column schemas are byte-identical to Jon's upstream parquet files, with
+one added column (`category`) on the markets table.
+
+If you use this dataset, please cite Jon Becker's original release.
+
+## How the Geopolitics subset was selected
+
+Each Polymarket market was classified into one of eight categories
+(Politics & Elections, Sports, Crypto & Web3, Finance & Economics,
+Geopolitics, Science & Technology, Entertainment & Culture, Other) by a
+two-tier text classifier:
+
+1. **Tier 1 — weighted keyword score.** For each candidate category, a
+   weighted hit count is computed over the market's `question` (weight
+   3), `slug` (weight 2), and `description` (weight 1). Categories use
+   topic-specific keyword lists (Geopolitics keywords include `war`,
+   `ukraine`, `russia`, `israel`, `gaza`, `nato`, `nuclear`, `ceasefire`,
+   `iran`, `north korea`, `taiwan`, `sanctions`, `geopolit`, `coup`,
+   `regime`, `summit`, `unsc`, `embargo`, `houthi`, `hezbollah`, `hamas`,
+   `border`, ...). The category with the highest score wins; ties are
+   broken by a fixed priority chain
+   (Politics > Geopolitics > Finance > Crypto > Sports > Science >
+   Entertainment > Other).
+2. **Tier 2 — TF-IDF + logistic-regression fallback.** Markets the
+   Tier-1 classifier sent to *Other* or scored below a single
+   low-weight keyword hit are re-classified by a TF-IDF + multinomial
+   logistic-regression model trained on a 640-row stratified
+   hand-labelled seed. Predictions with posterior < 0.4 remain in
+   *Other* to avoid mis-labelling low-confidence cases.
+
+Held-out weighted accuracy of the two-tier pipeline is 0.87 on the
+160-row eval split. The two-tier process drops the *Other* residual on
+the full 408 k-market sample from 44.9 % (Tier-1 only) to 9.3 %.
+
+This particular extract retains only markets whose final classifier
+assignment is **Geopolitics** (6,053 markets out of ~409 k).
+
+## Schemas
+
+### `markets/markets_*.parquet`
+
+| column | type | description |
+|---|---|---|
+| `id` | string | Polymarket market id (primary key). |
+| `condition_id` | string | Parent CTF condition id. |
+| `question` | string | The question prompt shown to users. |
+| `slug` | string | URL slug Polymarket uses to route this market. |
+| `outcomes` | string | JSON-encoded array of outcome labels, e.g. `["Yes","No"]`. |
+| `outcome_prices` | string | JSON-encoded array of final outcome prices. Resolution threshold > `0.99`. |
+| `clob_token_ids` | string | JSON-encoded array of CTF outcome-token ids (typically two: YES and NO). Use these to filter `trades/`. |
+| `volume` | double | Cumulative traded volume in USDC. |
+| `liquidity` | double | Snapshot of market depth. |
+| `active` | bool | Whether the market is currently open. |
+| `closed` | bool | Whether trading has ended. |
+| `end_date` | timestamp | Wall-clock end-of-life. |
+| `created_at` | timestamp | Wall-clock creation time. |
+| `market_maker_address` | string | Legacy FPMM contract address; NULL for CTF-Exchange markets (all Geopolitics markets here are CTF). |
+| `_fetched_at` | timestamp | Snapshot time at which this metadata row was scraped. |
+| `category` | string | **New column.** Always `"Geopolitics"` for this extract. |
+
+### `trades/trades_*.parquet`
+
+Each row is one on-chain trade fill. Unique key is
+`(transaction_hash, log_index)` — `transaction_hash` repeats because
+Polymarket uses batch fill.
+
+| column | type | description |
+|---|---|---|
+| `block_number` | int64 | Polygon block in which the trade was mined. |
+| `transaction_hash` | string | Polygon transaction hash. |
+| `log_index` | int32 | Position within the transaction's event log. |
+| `order_hash` | string | EIP-712 hash of the order. |
+| `maker` | string | Polygon address of the order maker (resting side). |
+| `taker` | string | Polygon address of the order taker (aggressing side). |
+| `maker_asset_id` | string | Asset id the maker delivered. `'0'` = USDC. |
+| `taker_asset_id` | string | Asset id the taker delivered. Mirror of `maker_asset_id`. |
+| `maker_amount` | int64 | Atomic units; divide by `1e6` for USDC or token units. |
+| `taker_amount` | int64 | Atomic units. |
+| `fee` | int64 | Polymarket fee in USDC atomic units. |
+| `timestamp` | timestamp | **Always NULL.** Join to `blocks/` to recover wall-clock. |
+| `_fetched_at` | timestamp | Scrape time. |
+| `_contract` | string | `"CTF Exchange"` or `"NegRisk CTF Exchange"`. |
+
+### `blocks/blocks_*.parquet`
+
+| column | type | description |
+|---|---|---|
+| `block_number` | int64 | Polygon block number. |
+| `timestamp` | string (ISO-8601) | Block timestamp, cast to `POSIXct` for arithmetic. |
+
+### Computing trade price and signed direction
+
+```
+price_cents = 100 * usdc_amount / token_amount
+            = 100 * (maker_amount if maker_asset_id == '0' else taker_amount)
+                  / (taker_amount if maker_asset_id == '0' else maker_amount)
+
+# Taker direction: positive = taker bought outcome tokens
+if maker_asset_id == '0':
+    taker_signed_size = -taker_amount / 1e6   # taker delivered tokens; sold
+else:
+    taker_signed_size = +maker_amount / 1e6   # taker delivered USDC; bought
+```
+
+## Quick start in R
+
+```r
+install.packages(c("arrow", "dplyr", "jsonlite", "duckdb"))
+library(arrow)
+library(dplyr)
+library(jsonlite)
+
+# Load the chunked parquet files transparently.
+markets <- open_dataset("polymarket/markets/")
+trades  <- open_dataset("polymarket/trades/")
+blocks  <- open_dataset("polymarket/blocks/")
+
+# Pick the largest-volume market.
+m <- markets %>%
+  arrange(desc(volume)) %>%
+  head(1) %>%
+  collect()
+
+# Parse the token ids and pull the matching trades.
+token_ids <- jsonlite::fromJSON(m$clob_token_ids)
+market_trades <- trades %>%
+  filter(maker_asset_id %in% token_ids | taker_asset_id %in% token_ids) %>%
+  collect() %>%
+  arrange(block_number, log_index) %>%
+  left_join(collect(blocks), by = "block_number") %>%
+  mutate(timestamp = as.POSIXct(timestamp, tz = "UTC"))
+
+# Derive price and signed direction.
+market_trades <- market_trades %>%
+  mutate(
+    is_usdc_maker     = maker_asset_id == "0",
+    usdc_amount       = ifelse(is_usdc_maker, maker_amount, taker_amount),
+    token_amount      = ifelse(is_usdc_maker, taker_amount, maker_amount),
+    price_cents       = 100 * usdc_amount / token_amount,
+    taker_signed_size = ifelse(is_usdc_maker,
+                               -taker_amount / 1e6,
+                               +maker_amount / 1e6)
+  )
+```
+
+## Streaming via the HuggingFace `datasets` library (Python)
+
+```python
+from datasets import load_dataset
+
+markets = load_dataset("smf-ulm/polymarket-geopolitics", "markets", split="train")
+trades  = load_dataset("smf-ulm/polymarket-geopolitics", "trades",  split="train", streaming=True)
+blocks  = load_dataset("smf-ulm/polymarket-geopolitics", "blocks",  split="train")
+
+print(markets[0]["question"])
+```
+
+## License
+
+The derivative dataset is licensed
+[CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/). Please cite
+both this release and Jon Becker's upstream repository when using the
+data in published work.
+
+## Citation
+
+```bibtex
+@misc{polymarket_geopolitics_2026,
+  author       = {Padmaperuma, Oliver},
+  title        = {Polymarket Geopolitics: Markets, Trades, Blocks (derived subset)},
+  year         = {2026},
+  howpublished = {HuggingFace dataset, smf-ulm/polymarket-geopolitics},
+  note         = {Derived from Jon Becker's polymarket-data release at
+                  \url{https://github.com/jbecker19/polymarket-data}}
+}
+```