Datasetpublic

rsoft-latam/erc8004-simulated-agents

ERC-8004 Simulated Agents — labeled synthetic dataset (6,000 agents) ⚠️ This dataset is fully synthetic. No public labeled dataset of malicious ERC-8004 agents exists (the standard reached mainnet in 2026 and exposes no trust label), so this dataset simulates the feature distributions the three ERC-8004 registries would expose, for training/evaluating trustworthiness models. For real on-chain data see the companion Base mainnet census. Composition 6,000 agents, 1… See the full description on the dataset page: https://huggingface.co/datasets/rsoft-latam/erc8004-simulated-agents.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face

0likes9downloads

Dataset Card

ERC-8004 Simulated Agents — labeled synthetic dataset (6,000 agents)

⚠️ This dataset is fully synthetic. No public labeled dataset of malicious ERC-8004 agents exists (the standard reached mainnet in 2026 and exposes no trust label), so this dataset simulates the feature distributions the three ERC-8004 registries would expose, for training/evaluating trustworthiness models. For real on-chain data see the companion Base mainnet census.

Composition

6,000 agents, 1,870 malicious (31.2%):

`subtype`	rows	Description
`trust`	4,200	Legitimate agents, including atypical "cold start" profiles (and 70 sophisticated adversaries that mimic normal agents, labeled malicious)
`sybil`	720	Mass-created identities: young, low completeness, low issuer diversity
`launder`	600	Reputation laundering: bursts of low-value positive feedback concentrated in time
`collusion`	480	Mutual-feedback rings: high reciprocity and clustering coefficient

Class overlap is intentional (sophisticated adversaries + atypical legitimate agents) to keep the task realistically hard.

Columns

20 features + label (0 = trustworthy, 1 = malicious) + subtype (generation regime):

—Identity: identity_age_days, agent_card_completeness, linked_dids, identity_validated
—Reputation: rep_positive, rep_negative, rep_ratio, issuer_diversity, rep_recency_days, rep_velocity
—Validation (simulated — ValidationRegistry is not on Base mainnet): val_requested, val_passed, val_failed, validator_avg_reputation
—Behavior/temporal/graph: interaction_count, interaction_value_avg, temporal_regularity, graph_degree, graph_clustering_coef, reciprocity

Intended use

Benchmarking anomaly detectors and supervised baselines for agent-trust research (used to train/evaluate AgentTrust-8004: Isolation Forest AUC-ROC 0.869 label-free; XGBoost ceiling 0.959 with labels). Do not report metrics from this dataset as real-world performance.

Attribution (CC-BY-4.0)

Pari Cahuna, R. E. (2026). ERC-8004 Simulated Agents. Master's in Data Science & AI, UAGRM School of Engineering. https://huggingface.co/datasets/rsoft-latam/erc8004-simulated-agents