xylemgroup/hyperliquid-chain-captures

Hyperliquid Chain Captures

Date-partitioned captures of HyperLiquid chain state for research use. Each date partition contains both raw node outputs (replica command logs, hourly event dumps, L2 book snapshots) and processed Parquet files derived from those raw captures.

The dataset is forward-compatible for additional dates: each date lives under its own top-level directory.

Layout

<date>/                                 # e.g., 2026-03-01/
├── raw/                                # native node outputs, lz4-compressed where applicable
│   ├── replica_cmds/
│   │   └── <capture_timestamp>/
│   │       └── <date>/
│   │           └── <block_number>      # ~2-7 GB per file, block-range chunks
│   ├── asset_ctxs/
│   │   └── <date>.csv.lz4
│   ├── misc_events_by_block/
│   │   └── hourly/<date>/<hour>        # per-hour event bundle
│   └── market_data/
│       └── <date>/<window>/l2Book/<symbol>   # L2 book per symbol per window
└── processed/                          # derived Parquet (zstd-compressed)
    ├── replica_cmds/part-0000.zstd.parquet
    ├── asset_ctxs/part-0000.zstd.parquet
    ├── misc_events_by_block/part-{0000..N}.zstd.parquet
    └── market_data/part-0000.zstd.parquet

Each processed/ directory also contains a _manifest.json documenting the row count, schema fingerprint, and source raw files for the Parquet output.

Loading the processed data

import pandas as pd

# Replica commands for a given date
df = pd.read_parquet(
    "https://huggingface.co/datasets/xylemgroup/hyperliquid-chain-captures/resolve/main/2026-03-01/processed/replica_cmds/part-0000.zstd.parquet"
)

Or via datasets:

from datasets import load_dataset
ds = load_dataset(
    "xylemgroup/hyperliquid-chain-captures",
    data_files="2026-03-01/processed/replica_cmds/part-0000.zstd.parquet",
)

Available dates

Date	Raw size	Processed size	Block ranges covered (replica_cmds)
2026-03-01	~29.1 GB	~4.6 GB	909,310,000–909,330,000 and 909,615,000–909,630,000

Source

HyperLiquid chain replica state, captured via the official replica command logs. The raw replica_cmds files are the unmodified node-side output for the indicated block ranges. The processed Parquet files are deterministic derivations of the raw inputs — schemas and source-file lineage are recorded per-dataset in the corresponding _manifest.json.

License

Released under CC0 1.0 (public domain dedication). The underlying chain data is public; this dataset is the captured form of it. No attribution required.