software-ses/raven-dataset

Dataset Card for RAVEN: Analyzing Ethereum’s Reverted Transactions via Semantic Clustering of Failure Invariants

Dataset Description

This dataset comprises two collections (splits) of failed transactions on the Ethereum blockchain, annotated with extracted business‑logic invariants. The dataset was created within the research project titled HighGuard: Cross‑Chain Business Logic Monitoring of Smart Contracts, by Mojtaba Eshghie.

• Finetuning collection: ~100,000 failed Ethereum transactions annotated with 1,932 unique invariants.
• Evaluation collection: ~20,000 sampled failed transactions annotated with 727 unique invariants, used for clustering and categorization evaluation.

Each record corresponds to a failed transaction, along with metadata such as transaction hash, block number, sender/receiver, gas used/limit, failure message, and extracted invariant condition that caused the failure.

Key features

• Focused on business‐logic vulnerabilities, not only low‑level errors (e.g., out‑of‑gas) but semantic violations captured via invariants.
• Two distinct collections (finetuning + evaluation) for training and benchmarking.
• Designed for anomaly‑detection and classification tasks in the smart‑contract security domain.

Recommended uses

• Training supervised or unsupervised models to detect business‑logic failures in smart contracts.
• Clustering sampled invariants to categorize common failure types.
• Benchmarking research on smart‐contract verification, transaction analysis, and runtime monitoring.

Out‑of‑Scope uses

• This dataset is not suitable for general cryptocurrency transaction modelling (e.g., normal transfers), since only failed transactions are included.
• It is not a comprehensive dataset of all Ethereum transactions — only those with business‐logic failure annotations.

---

Dataset Structure

The dataset is provided as a DatasetDict with two splits/collections:

Split	Description	Approx. Size
finetuning	100 000 failed transactions annotated with 1 932 invariants	~100k rows
evaluation	20 000 failed transactions annotated with 727 invariants	~20k rows

Each record has the following columns:

• tx_hash (string): Transaction hash.
• block_number (int64): Block number in which the transaction was included.
• from_address (string): Sender Ethereum address.
• to_address (string): Receiver Ethereum address.
• gas_limit (int64): Gas limit specified for the transaction.
• gas_used (int64): Gas used by the transaction before failure.
• failure_message (string): The revert or failure message (if available).
• invariant_condition (string): A high‐level invariant representing the business‐logic violation.
• invariant_id (int64): An internal identifier for the extracted invariant cluster/category.
• timestamp (int64): Unix timestamp of the block (optional).

File format: The repository provides Parquet files for each split (finetuning.parquet, evaluation.parquet) and can be loaded via the datasets library as:

from datasets import load_dataset
ds = load_dataset("MojtabaEshghie/raven‑dataset", split="finetuning")

---

Dataset Creation

Curation Rationale

Business‐logic failures in smart contracts are harder to detect than low‐level exceptions (e.g., out‑of‑gas) but are critically important for security. The goal of this dataset is to provide a curated collection of failed transactions with extracted invariants to enable anomaly detection, clustering, and classification research in the smart‐contract domain.

Data Processing

• Filtering of failed transactions with revert/failure messages.
• Extraction of business‑logic invariants via the tool SoliDiffy and other analysis pipelines.
• Deduplication of similar invariant texts and clustering of invariants to create invariant_id.
• Serialization into Parquet format; conversion to Arrow format by the datasets library during upload.

Who/When/Where

• Curated by: Mojtaba Eshghie
• Affiliation: KTH Royal Institute of Technology, Umeå University.
• Date: Nov 2025

---

Considerations for Using the Data

Limitations

• Bias toward failures only: The dataset contains only failed transactions, so models trained on it might not generalize to normal transactions.
• Time cutoff: Transactions are up to a certain block number.

Ethical and Privacy Considerations

• The data is sourced from a public blockchain (Ethereum), so transaction data is publicly available.
• Addresses are included (sender/receiver), which are pseudonymous but publicly traceable; users should be aware of potential linking to identities through external sources.
• Use responsibly: do not attempt to de‑anonymize addresses or misuse user data.

Recommendations

• If using for supervised classification, consider balancing via sampling or weighting due to potentially unbalanced invariant categories.
• For anomaly detection, consider using the finetuning split for training and evaluation for benchmarking.
• Always cite the dataset and the associated paper when using it in publications.

---

Citation

tbd