Hugging Face's logo

Datasets:
ironflow
/
hyperliquid-liquidations

Tasks:
Time Series Forecasting
Modalities:
Tabular
Text
Formats:
parquet
Size:
10M - 100M
Tags:
hyperliquid
crypto
perpetual-futures
defi
liquidations
finance
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
hyperliquid-liquidations / README.md
ironflow's picture
ironflow
v3.1: 2026-05, 06 and 10 re-exported after repairs (refilled hours, 2,545 double-counted rows removed)
65bad6c verified
preview code
|
Raw History Blame Contribute Delete
8.33 kB
metadata 
license: cc-by-4.0
pretty_name: Hyperliquid Liquidations
task_categories:
  - time-series-forecasting
tags:
  - hyperliquid
  - crypto
  - perpetual-futures
  - defi
  - liquidations
  - finance
  - trading
  - time-series
size_categories:
  - 1M<n<10M
configs:
  - config_name: events
    data_files:
      - split: train
        path: data/events/*.parquet
    default: true
  - config_name: fills
    data_files:
      - split: train
        path: data/fills/*.parquet

Hyperliquid Liquidations

Every liquidation on Hyperliquid perpetuals from 25 May 2025 to 31 October 2026: 2,849,452 liquidated positions across 468 markets and 506,975 wallets, $58.7 billion in liquidated notional. Native perps and HIP-3 builder markets (stocks, commodities, indices) are both included. The busiest day is the 10 October 2025 crash: 109,790 liquidations, $7.2 billion.

Most of this history is no longer available from Hyperliquid's own API, which returns only a wallet's most recent 2,000 fills (10,000 by time range). The wallet behind the two largest liquidations of 10 October 2025 ($204.7M ETH and $197.2M BTC, $1.3 billion across four markets in under three minutes) reaches back only to August 2026 there.

The data comes from Ironflow, which runs its own Hyperliquid nodes and reads the liquidation field Hyperliquid puts on every liquidation fill. It is not a heuristic: each row is a fill Hyperliquid itself marked as a liquidation. Months before November 2025 were replayed from Hyperliquid's node archive on S3 and checked hour by hour against the archive files.

Get live data

Liquidations as they happen, on every Hyperliquid perp: ironflow.sh/liquidations. Other live Hyperliquid data is on Ironflow.

For data newer than this snapshot, the same rows are on Ironflow's API: GET /v1/liquidations for history and GET /v1/liquidations/stream for a live push. No key is needed for the last 24 hours; a free key gives 30 days.

Source and updates

  • Source: liquidation fills recorded by Ironflow's own Hyperliquid nodes. Months before November 2025 were replayed from Hyperliquid's node archive on S3, and hours that node incidents had left short were refilled from the same archive (v2).
  • Checks: replayed months checked hour by hour against the archive files; trade counts reconciled with Hyperliquid's own candles; randomly sampled events match Hyperliquid's userFillsByTime exactly (see Changes).
  • Updates: refreshed monthly, after each UTC month closes. For anything newer, use the live page or the API.
  • Coverage: 25 May 2025 to 31 October 2026. Native perps and HIP-3 perps, plus a few spot fills (see Notes).

Files

Two tables, one Parquet file per month (UTC):

Config Path One row per
events (default) data/events/YYYY-MM.parquet liquidated position: all fills of one wallet in one market in one block
fills data/fills/YYYY-MM.parquet liquidation fill, as Hyperliquid recorded it

A liquidation usually closes in a few fills (2.8 on average); a large one can take hundreds. Use events to count and size liquidations, fills for execution detail.

events

Column Type Meaning
timestamp timestamp (ms, UTC) time of the first fill
market string Ironflow market symbol, e.g. BTC-PERP, xyz:NVDA-PERP
issuer string HIP-3 builder (xyz, flx, ...); empty for native perps
wallet string the liquidated wallet
position_side string long or short: the position that was liquidated
size float total size closed, in base units
notional_usd float sum of price x size over the fills
avg_price float size-weighted average fill price
fills int number of fills
block int Hyperliquid block; null before 27 July 2025 (see Notes)
tx_hash string hash of one of the fills

fills

Column Type Meaning
timestamp timestamp (ms, UTC) fill time
market, issuer, base_market string market symbol, HIP-3 builder, symbol without the builder prefix
wallet string the liquidated wallet
side string side of the liquidated wallet's closing fill: sell closes a long, buy closes a short
price, size, notional_usd float fill price, size and price x size
block, tx_hash int, string Hyperliquid block (null before 27 July 2025) and fill hash

Use 

import pandas as pd

events = pd.read_parquet("hf://datasets/ironflow/hyperliquid-liquidations/data/events/")
daily = events.groupby([events.timestamp.dt.date, "position_side"]).notional_usd.sum().unstack()
print(daily.tail())

-- DuckDB: the 10 largest liquidations
SELECT timestamp, market, position_side, round(notional_usd) AS usd, wallet
FROM 'data/events/*.parquet'
ORDER BY notional_usd DESC
LIMIT 10;

Notes

  • Only the liquidated wallet's fills are included. The counterparty's fill carries the same liquidation field on Hyperliquid; it is dropped so each liquidation counts once.
  • A few spot markets appear too (104 fills, about $0.75M): spot positions closed by portfolio-margin liquidations.
  • Prices and sizes are floats as Hyperliquid reports them; notional is in the market's quote asset, USD-denominated (USDC) for every market here.
  • Wallet addresses are public on-chain data.
  • Before 27 July 2025 08:50 UTC Hyperliquid's archive stores fills without block numbers, so block is null there and an event groups one wallet's fills in one market at one block timestamp (every fill of a block shares it). The archive has no fills between 08:45:48 and 08:50:10 UTC that day, and the first file starts on 25 May 2025 at 14:00 UTC.
  • Wallet addresses in this dataset come from public Hyperliquid data. To object to, or ask us to remove, a wallet address, write to hello@ironflow.sh. Versions already published cannot be recalled; future versions leave out any address we are asked to remove.

Changes

  • v3.1, 7 November 2026. Three months re-exported after repairs on our side; the other 15 are unchanged. 2026-05 and 2026-06 gain the first days of 14 HIP-3 listings and hours that node restarts had left short, refilled from Hyperliquid's S3 archive (+1,257 and +2,945 fills). They also lose 2,545 rows that counted a liquidation twice: 2,411 in 2026-06 that recorded the counterparty's side under the liquidated wallet, and 134 stored twice (33 in May, 101 in June); each kind was checked against Hyperliquid's userFillsByTime. 2026-10 gains 16 fills lost on node restarts. 2,849,452 events in all (+1,039). Sampled new events match userFillsByTime exactly (11 of 11 Hyperliquid still serves).
  • v3, 1 November 2026. Starts on 25 May 2025 instead of 1 November 2025 (Hyperliquid's archive reaches back that far) and runs to 31 October 2026: 2,848,413 events against v2's 2,167,823, including the 10 October 2025 crash. The months already in v2 are unchanged, row for row. Daily trade counts of the replayed months match Hyperliquid's own candles within 0.5% (10 October 2025 has 0.4 to 1.5% more, all in the 20:00 to 24:00 UTC window of the cascade). 12 randomly sampled events from the new months, two per month, match Hyperliquid's userFillsByTime exactly.
  • v2, 31 October 2026. We reconciled our fills against Hyperliquid's own candles, hour by hour over the 12 months, and refilled the hours that node incidents had left short (largest on 29 April, 18 to 22 May and 1 to 7 June 2026) from Hyperliquid's S3 archive. v2 has 37,269 more liquidations than v1 (2,167,823 against 2,130,554); 29 April 2026 went from 888 to 6,088 events. 24 randomly sampled events, two per month, match Hyperliquid's userFillsByTime exactly. The window now starts on 1 November 2025.
  • v1, 30 October 2026. First release.

License and citation

CC BY 4.0. Credit "Ironflow (ironflow.sh)" when you use or share the data.