Intent2Tx/web3_intents_to_ethereum_transactions

Natural Language Web3 Intents to Ethereum Transactions

This dataset contains instruction-following examples for mapping natural language Web3 intents to executable Ethereum transaction plans.

It is organized into two JSONL splits:

• single_step: one user intent mapped to one on-chain action (hf_single_step_intent2tx.jsonl)
• multi_step: one user intent mapped to an ordered list of on-chain actions (hf_multi_steps_intent2tx.jsonl)

Dataset Summary

• Total examples: 31,496
• single_step: 29,921 examples
• multi_step: 1,575 examples
• Language: English (en)
• Domain: Ethereum / Web3 intent-centric planning

Data Instances

Single-step split (single_step)

Each record includes:

• instruction (string): task instruction template
• input (string): user intent text (prefixed with "User intent:\n")
• output (string): JSON string for a single action plan
• contract (string): target contract name
• function (string): target function name
• primary_category (string): coarse intent category
• sub_category (string): fine-grained intent category
• tx_hash (string): Ethereum transaction hash
• metadata (object): additional on-chain context (e.g., block time, selector, calldata, sender, value, gas used)

Multi-step split (multi_step)

Each record includes:

• instruction (string): task instruction template
• input (string): user intent text (prefixed with "User intent:\n")
• output (list[object]): ordered action plan with one or more on-chain actions

Each action in output follows:

• contract (string)
• contract_address (string)
• function (string)
• params (object): function parameter schema and values
• value (float): ETH value sent with this action

Intended Use

This dataset is intended for:

• Training and evaluating LLMs for intent-to-transaction planning
• Benchmarking single-action and multi-action decomposition capabilities
• Research on structured generation for Ethereum execution workflows

Loading the Dataset

You can load this dataset from local files or directly from the Hugging Face Hub.

from datasets import load_dataset

# If files are local
ds = load_dataset(
    "json",
    data_files={
        "single_step": "hf_single_step_intent2tx.jsonl",
        "multi_step": "hf_multi_steps_intent2tx.jsonl",
    },
)

print(ds)
print(ds["single_step"][0].keys())

Limitations

• Outputs are automatically derived from historical on-chain activity and may contain noise.
• The dataset reflects Ethereum behavior in a specific collection period, and may not represent future protocol changes.
• Generated action plans should be validated before real-world execution.

Ethical and Safety Considerations

• This dataset is for research and benchmarking purposes.
• Do not use model outputs directly for financial decisions or production fund management without additional safeguards and simulation/testing.