Thor25: THORChain Cross-Chain Data

This repository contains the Thor25 dataset released with LOCARD: An Agentic Framework for Blockchain Forensics, published at IEEE ICBC 2026, Brisbane, Australia, June 1-5, 2026.

Thor25 supports research on agentic blockchain forensics, cross-chain transaction tracing, and evidence-grounded tool use by LLM agents. It does not map cleanly to conventional NLP task categories such as question answering or text generation: solving each benchmark case requires an agent to iteratively interact with blockchain data and investigative tools in a real blockchain environment.

• Paper DOI: https://doi.org/10.1109/ICBC67748.2026.11575479
• Hugging Face paper page: https://huggingface.co/papers/2604.04211
• arXiv: https://arxiv.org/abs/2604.04211
• Hugging Face dataset: https://huggingface.co/datasets/xhyumiracle/thor25
• LOCARD code: https://github.com/xhyumiracle/locard

About THORChain

THORChain is a decentralized cross-chain liquidity protocol built as an independent Layer 1 blockchain using Cosmos SDK. It enables native asset swaps across different blockchains without wrapped tokens or centralized custodians. The THORChain network produces blocks approximately every 6 seconds (docs).

Data Selection

THORChain supports cross-chain swaps across many blockchains (BTC, ETH, BSC, AVAX, DOGE, LTC, BCH, etc.) and various assets (native coins, ERC-20 tokens, BEP-20 tokens, etc.).

This dataset focuses on:

• Chains: BTC, ETH, DOGE, LTC (4 common chains)
• Assets: Native assets only (BTC, ETH, DOGE, LTC)
• Filter: type=swap and status=success records only

Dataset Naming Convention

Format: thorchain-2025-<condition1>-<condition2> (Short: Thor25<C1><C2>)

Amount: L (Low) | H (High) | X (eXtra) Time: F (Fast) | S (Slow)

Current datasets:

• thorchain-2025 (Thor25): Full 2025 dataset (151,461 standard 1-in-1-out swaps)
• thorchain-2025-high-fast (Thor25HF): High amount (0.09 BTC / 1.9 ETH / 2.5 LTC / 1k DOGE) + Fast completion (≤30min, 20,235 records)
• thorchain-2025-high-fast-mini (Thor25HF-mini): Mini test set sampled from Thor25HF (1,200 records, 100 per pair)
• thorchain-2025-multi (Thor25M): Multi-output swaps (166 records, currently not used for queries)

Hugging Face Usage

from datasets import load_dataset

dataset = load_dataset("xhyumiracle/thor25", "thor25hf-mini")
print(dataset["test"][0])

Load the full dataset:

from datasets import load_dataset

dataset = load_dataset("xhyumiracle/thor25", "thor25")

Quick Start

Query files are not included in the repository (they're generated from the source data). Generate them locally:

# Generate queries for mini test set (1,200 queries, recommended for first try)
uv run python script/process/gen_query.py --batch --input-dir data/thorchain-2025-high-fast-mini --output-dir queries/thorchain-2025-high-fast-mini

# Generate queries for high-fast dataset (~20.2k queries)
uv run python script/process/gen_query.py --batch --input-dir data/thorchain-2025-high-fast --output-dir queries/thorchain-2025-high-fast

# Generate queries for full dataset (~151k queries)
uv run python script/process/gen_query.py --batch --input-dir data/thorchain-2025 --output-dir queries/thorchain-2025

Data Fields

Each record contains:

• idx: dataset-local sequential index.
• id: hash-based stable identifier derived from transaction entries.
• timestamp: THORChain action timestamp in Unix nanoseconds, UTC.
• type: THORChain action type. This release keeps swap records.
• status: THORChain action status. This release keeps success records.
• in: list of source transaction entries.
• out: list of destination transaction entries.

Each entry in in and out includes:

• chain: blockchain name, one of BTC, ETH, DOGE, or LTC.
• asset: native asset name.
• txID: blockchain transaction hash.
• address: sender or receiver address.
• amount: asset amount in THORChain Midgard base units.
• thorchainHeight: THORChain block height.

Data Characteristics

Overview

• Time Range: 2025-01-01 00:00:00 UTC ~ 2025-12-31 23:59:59 UTC (full year)
• Total Records: 151,461 successful swaps across 12 pair files
• Pairs: BTC<>ETH, BTC<>DOGE, BTC<>LTC, ETH<>DOGE, ETH<>LTC, DOGE<>LTC (both directions)

Amount Unit Normalization

THORChain Midgard API normalizes all asset amounts to 1e8 base units (similar to Bitcoin's satoshi), regardless of the native blockchain's decimal precision (Midgard docs):

• BTC (native 1e8 satoshi) → preserved as 1e8
• ETH (native 1e18 wei) → shortened to 1e8
• DOGE (native 1e8) → preserved as 1e8
• LTC (native 1e8 litoshi) → preserved as 1e8

This means: 1 BTC = 1 ETH = 1 DOGE = 1 LTC = 100,000,000 units in the data.

Timestamp Format

All timestamps in Midgard API are Unix timestamps in UTC timezone, with the date field in nanoseconds (API spec).

Record Indexing

Each record has two identifiers: idx (dataset-local sequential index starting from 0) and id (hash-based stable identifier across all datasets).

Transaction Count by Direction

Full Dataset (151,461 records):

• BTC→ETH: 50,597 records (33.4%)
• ETH→BTC: 37,087 records (24.5%)
• LTC→BTC: 18,090 records (11.9%)
• BTC→LTC: 11,256 records (7.4%)
• LTC→ETH: 9,503 records (6.3%)
• ETH→LTC: 7,819 records (5.2%)
• DOGE→BTC: 5,149 records (3.4%)
• DOGE→ETH: 3,782 records (2.5%)
• BTC→DOGE: 3,482 records (2.3%)
• ETH→DOGE: 2,560 records (1.7%)
• DOGE→LTC: 1,016 records (0.7%)
• LTC→DOGE: 1,120 records (0.7%)

High-Fast Dataset (20,235 records):

• BTC→ETH: 6,511 records (32.2%)
• ETH→BTC: 4,928 records (24.4%)
• LTC→BTC: 3,234 records (16.0%)
• LTC→ETH: 2,038 records (10.1%)
• DOGE→BTC: 1,469 records (7.3%)
• DOGE→ETH: 739 records (3.7%)
• BTC→LTC: 639 records (3.2%)
• ETH→LTC: 246 records (1.2%)
• DOGE→LTC: 117 records (0.6%)
• BTC→DOGE: 107 records (0.5%)
• LTC→DOGE: 104 records (0.5%)
• ETH→DOGE: 103 records (0.5%)

All pairs filtered with: source tx amount thresholds (≥0.09 BTC / ≥1.9 ETH / ≥2.5 LTC / ≥1000 DOGE) + time constraint (≤30min).

Daily Distribution

📊 LTC Pairs (Click to expand)

Amount Distribution

• Amount ranges span multiple orders of magnitude (10³ to 10¹⁰)
• Each asset shows distinct distribution patterns
• IN (solid line in CDF) vs OUT (dashed line in CDF) amounts show swap behavior

📊 LTC Pairs (Click to expand)

Height Diff (Swap Completion Time)

• Most swaps complete quickly: median 6-26 thorchain blocks depending on pair
• ~80-90% complete within 100 blocks
• ~99%+ complete within 1000 blocks
• Outliers exist up to 24k blocks (ETH→BTC)

📊 LTC Pairs (Click to expand)

Time Diff (Blockchain Timestamps)

• Time difference calculated from actual blockchain transaction timestamps
• Provides real-world completion time in seconds (vs THORChain blocks)
• Shows directional asymmetry: BTC→DOGE (177s) vs DOGE→BTC (762s)
• Most swaps complete within 500 seconds

📊 LTC Pairs (Click to expand)

Time-Amount Relationship

• Visualization of time difference vs transaction amount across all pairs
• High-Fast (HF) dataset thresholds shown as reference lines
• Helps understand the characteristics of filtered datasets

Traffic Spikes

• 2025-02-2203-03: Major spike in ETH→BTC (10 days, tx count surged from ~100/day to 1000-2000/day, daily amount jumped from ~100 ETH to 20,000-85,000 ETH), related to Bybit hack fund flows
• 2025-03-14~15: Abnormal spike in BTC<>ETH (1800+ tx/day, height diff up to 5000+ blocks)
• 2025-06: Another spike (~1100 tx/day, height diff up to 3000+ blocks)

Scripts

crawl/ - Fetching Raw Data and Reprocessing

fetch_swaps.py

Midgard API crawler (backwards by timestamp) for THORChain swap actions.

• --min-ts: Lower bound Unix timestamp, stop when reaching this
• --max-ts: Upper bound Unix timestamp, start from here
• --fresh: Start a new crawl from scratch
• --resume: Continue from last saved state

# Fresh crawl
uv run python script/crawl/fetch_swaps.py --outdir raw --min-ts 1735689600 --fresh

# Resume
uv run python script/crawl/fetch_swaps.py --outdir raw --min-ts 1735689600 --resume

wash.py

Transform raw data to cleaned format. Filters status != 'success' records and removes THOR.* assets (e.g., THOR.RUNE affiliate fees) from outputs to keep only the actual swap assets.

Record ID: id = SHA-256("\n".join(sorted(entries)) + "\n{type}|{status}") where each entry = "{direction}|{chain}|{asset}|{address}|{txID}" (auto-deduplicated). This derived ID appears as query_id in generated query files.

uv run python script/crawl/wash.py --indir raw/data --outdir data/thorchain-2025

process/ - Data Processing Pipeline

filter_data.py

Filter swap data by amount and time thresholds to create high-quality datasets.

# Custom amount thresholds (see script header for examples)
uv run python script/process/filter_data.py --src-amount-gte-BTC 0.09 --src-amount-gte-ETH 1.9 --src-amount-gte-LTC 2.5 --src-amount-gte-DOGE 1000

sample_mini.py

Sample mini dataset from high-fast data for testing.

uv run python script/process/sample_mini.py

gen_query.py

Generate YAML batch query files from jsonl data.

# Generate from a single jsonl file
uv run python script/process/gen_query.py --input ../../data/BTC-DOGE.jsonl --output ../../queries/BTC-DOGE.yaml

# Generate from all jsonl files (batch mode)
uv run python script/process/gen_query.py --batch --input-dir ../../data --output-dir ../../queries

# # Optional: Add timestamp_delta metadata (requires blockchain_txs/ directory from enrich/)
# uv run python script/process/gen_query.py --batch --input-dir ../../data --output-dir ../../queries --blockchain-txs-dir ../../blockchain_txs

The generated YAML files can be used with the system:

cd <project_root>
python -m src.main --batch data/thorchain/queries/BTC-DOGE.yaml

analyze/ - Tools for Validate and Analyzing Data

validate.py

Check for duplicate records by ID.

uv run python script/analyze/validate.py

stats.py

Compute per-pair statistics (amounts, height diff, timestamps).

uv run python script/analyze/stats.py

plot.py

Plot amount & height diff vs timestamp.

uv run python script/analyze/plot.py

Output: png/*.png

identify_slow_swaps.py

Identify swaps with abnormally long completion times (for debugging/analysis).

• -t: Height diff threshold (records with diff > threshold)
• -s: Start date (YYYY-MM-DD)
• -e: End date (YYYY-MM-DD)

# Basic
uv run python script/analyze/identify_slow_swaps.py -t 5000

# With date range
uv run python script/analyze/identify_slow_swaps.py -t 2000 -s 2025-03-01 -e 2025-03-31

# Export JSON
uv run python script/analyze/identify_slow_swaps.py -t 5000 -o results.json

Intended Uses

This dataset is intended for research on blockchain forensics and cross-chain transaction tracing, evaluation of agentic AI systems that perform evidence-grounded investigation, benchmarking retrieval/candidate-generation/ranking/forensic-reasoning workflows, and reproducibility of LOCARD experiments.

Limitations

Thor25 is not a complete dataset of all THORChain activity. It focuses on selected native assets, successful swaps, and benchmark-friendly record types. Blockchain attribution is inherently uncertain, and benchmark results should not be interpreted as legal, compliance, or financial conclusions.

The dataset contains public blockchain addresses and transaction hashes. Users should avoid attempting to deanonymize individuals or use the dataset for harmful surveillance.

Citation

@inproceedings{YuK26,
  title     = {LOCARD: An Agentic Framework for Blockchain Forensics},
  author    = {Xiaohang Yu and William Knottenbelt},
  booktitle = {2026 IEEE International Conference on Blockchain and Cryptocurrency (ICBC)},
  pages     = {1--9},
  publisher = {IEEE},
  year      = {2026},
  doi       = {10.1109/ICBC67748.2026.11575479},
  url       = {https://doi.org/10.1109/ICBC67748.2026.11575479}
}

License

Thor25 is released under the Creative Commons Attribution 4.0 International license (cc-by-4.0).