ProbioticFarmer/hyperliquid-augmented-copytrading

Hyperliquid L1 Consensus & Microstructure Research Archive

Georgia Tech CSE6242 (Data & Visual Analytics) — Semester Project

An institutional-grade, multi-stream market microstructure and trader behavior archive covering the complete trading history of the Hyperliquid L1 decentralized exchange.

Published under the MIT license to allow academic researchers, peer graders, and students to reproduce all trader rankings, conformal predatory swindle detectors, and execution latency models with zero egress fees.

---

⚡ Fast Streaming Guide (Zero-Download Footprint)

The repository is organized into distinct Configurations (Subsets) according to Hugging Face best practices. You can stream individual records on-the-fly using HTTP range requests without downloading the 249+ GB corpus.

1. Python datasets Streaming Quickstart

from datasets import load_dataset

# Stream perpetual contract trade executions
ds_perp = load_dataset(
    "ProbioticFarmer/hyperliquid-augmented-copytrading",
    name="perp_fills",
    split="train",
    streaming=True
)

for row in ds_perp:
    print(f"Coin: {row['coin']} | Price: {row['price']} | Size: {row['size']} | Side: {row['side']} | User: {row['user']}")
    break

2. Stream Daily Account Equity Balances

# Stream daily account equity for 460k wallets (survivor bias mitigation)
ds_accounts = load_dataset(
    "ProbioticFarmer/hyperliquid-augmented-copytrading",
    name="account_values",
    split="train",
    streaming=True
)

for row in ds_accounts:
    print(f"User: {row['user']} | Total Equity: ${row['account_value']:,.2f}")
    break

3. Stream Column Subsets (Network-Bandwidth Optimization)

Because Parquet is columnar, you can stream only the specific columns your model needs, bypassing all unneeded data:

# Stream only timestamps, coins, prices, and sizes
ds_filtered = load_dataset(
    "ProbioticFarmer/hyperliquid-augmented-copytrading",
    name="perp_fills",
    split="train",
    streaming=True,
    columns=["coin", "price", "size", "side"]
)

4. Serverless SQL Queries via DuckDB (Direct Remote Range Requests)

import duckdb

con = duckdb.connect()

# Query trade executions directly across Hugging Face without downloading files
query = \"\"\"
SELECT 
    coin,
    side,
    price,
    size,
    user
FROM 'hf://datasets/ProbioticFarmer/hyperliquid-augmented-copytrading/perp_fills_all/date=2026-10-09/fills.parquet'
WHERE coin = 'BTC'
LIMIT 10;
\"\"\"
print(con.execute(query).df())

---

📦 Dataset Inventory & Configurations

Configuration (name=)	File Count	Size (GB)	Partitioning	Description
perp_fills	443	140.31 GB	date=YYYY-MM-DD/	Every perpetual fill from DEX genesis across all crypto + synthetic tokens
spot_fills	443	16.33 GB	date=YYYY-MM-DD/	All spot market trade executions across all token pairs
account_values	391	18.12 GB	date=YYYY-MM-DD/	Daily equity for all ~460k accounts from inception (mitigates survivor bias)
candles_1s	440	11.64 GB	date=YYYY-MM-DD/	High-frequency 1-second OHLCV candles across all perpetual markets
orderbook_1m	55,579	13.67 GB	date=YYYY-MM-DD/coin.parquet	Full-DEX 1-minute resting bid/ask depth across all 150+ perpetual markets
snapshots_spot	391	32.63 GB	date=YYYY-MM-DD/	Daily spot token balances and wallet holdings across all accounts
snapshots_perp	391	5.65 GB	date=YYYY-MM-DD/	Daily perpetual leverage, margin tier, and open positions per wallet
liquidations	443	0.32 GB	date=YYYY-MM-DD/	Event-level cascade records, liquidation sizes, and price dislocation ticks
outcome_fills	330	1.07 GB	Hierarchical	Binary prediction and outcome event contract executions
asset_ctxs	1,242	9.51 GB	Flat daily CSV.lz4	3.4 years of hourly funding rates, open interest, and oracle mark prices
TOTAL CORPUS	60,093	249.24 GB	—	Complete DEX Market Microstructure & Trader Behavior Archive

---

🔬 Supported Research Modules

1. Augmented Copy-Trading Engine: Real-time trader profiling, Sortino/Sharpe re-ranking, and survivorship-bias filtering across 460k wallets.
2. Conformal Microstructure Shield: Mondrian conformal prediction model (CatBoost base learner) filtering predatory order-book sweeps ("scamwicks").
3. Execution Latency Backtest: Millisecond-level queue position and taker slippage simulation.

---

⚖️ Citation & Acknowledgments

• Primary Source: Hyperliquid L1 Consensus Replay & Hydromancer Reservoir (s3://hydromancer-reservoir).
• Academic Course: Georgia Tech CSE6242 Data and Visual Analytics (Team 176).