Pseudonymised per-fill markout and order-book dataset for a counterparty adverse-selection study on the Hyperliquid perpetual DEX (BTC/ETH, December 2025)

Per-fill markout and order-book data used by the reproducibility notebook of an MScFE 690 capstone study of counterparty-resolved adverse selection, that is toxic order flow, on the Hyperliquid perpetual decentralized exchange, on BTC and ETH perpetuals over 4 to 13 December 2025. The fill table holds 2,480,267 rows, one per collapsed taker trade, across 45,849 taker and 9,367 maker pseudonyms. Each row carries the block time, the entry price, the notional, the taker side, the signed pre-trade inventory, whether the fill opens a position or closes one, signed markouts at 1, 10 and 60 seconds on the trade print, the realised profit the fill closes out, and integer counterparty and maker codes. The order-book series sit beside it on the book clock, un-resampled, at 1,572,177 snapshots for BTC and 1,572,228 for ETH. Each snapshot carries the mid, the microprice, the spread and the queue imbalance, and with them the feature block the identity-against-book comparison is measured on: order-flow imbalance, a book-based VPIN, queue imbalance at one and at ten levels, and bid and ask depth within 5 and within 10 basis points. The notebook rebuilds the mid markouts and the book state at each fill from these rather than reading either one precomputed, so it performs the as-of join itself and a reader can test the instant the book was read at. On identifiers. Counterparty and maker wallets are replaced by contiguous integer codes and the address map is not published, so these files carry no wallet address. That makes the codes pseudonyms and not anonymous labels, and the distinction is stated here because we measured it: the source tape this dataset derives from is itself public and carries the addresses, so joining a row back to it on coin, block time, price and size recovers the address for the large majority of codes. Nothing here is disclosed that the source deposit does not already disclose. Treat the codes as a convenience for analysis and not as a privacy control. Derived from the public Hyperliquid Level-4 order-book archive, Zenodo 10.5281/zenodo.18184441 (Albers, Cucuringu, Howison and Shestopaloff, "An Open Book", 2026), CC-BY 4.0. Files. Three, and the notebook reads all three: notebook_fills.npz, the per-fill table, 51 MB; notebook_book_BTC.npz and notebook_book_ETH.npz, the order-book series, 69 MB and 70 MB. The order-book files were called mid_series_BTC.npz and mid_series_ETH.npz in version 1 and hold eight more columns here, so they carry a new name rather than the old one over different contents. Version 1 also held two files the notebook no longer reads, notebook_data.npz and notebook_expected.json, and neither is carried forward. Version 1 stays reachable at its own address, so a citation made against it still resolves. The reported figures the notebook checks itself against now travel inside the notebook, so that a figure and the code that checks it cannot be of different ages.