HEART: Reinforcement Learning Dataset with Assistive and Adversarial Cues

A dataset for evaluating reinforcement learning agents, created by HPAI-BSC and last updated on 2026-05-04. It contains multiple splits, including a baseline with no injected cues and four cue-based splits. Each cue type is instantiated in both assistive (consistent with ground truth) and adversarial (supporting an incorrect option) variants.

Use Cases

• Benchmarking agent robustness based on the presence of injected environmental cues.
• Studying the impact of assistive cues on agent learning efficiency.
• Evaluating agent vulnerability to adversarial cues designed to mislead decision-making.
• Analyzing performance differences across multiple dataset splits with varying cue types.

Strengths

• Includes a controlled baseline split for performance comparison.
• Provides four distinct cue-based splits for varied testing scenarios.
• Each cue type has both assistive and adversarial variants, enabling nuanced evaluation.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

HPAI-BSC

Freshness

Last updated 2026-05-04 09:15:36; freshness should be verified.