106 million tokens of synthetic documents generated for evaluating AI models. Documents were created using GPT-4.1 for brainstorming and GPT-5 for final generation, based on the pipeline from Wang et al. (2025). The dataset is organized by seven evaluation traits, including Verifiable Structure, Placeholder, and Ethical Dilemma.
Use Cases
- Benchmarking model responses to Harmful Requests based on the described trait.
- Evaluating AI reasoning in scenarios with Conflicting Goals as described.
- Testing model consistency under an Inconsistent Environment as mentioned.
- Assessing ethical decision-making in Ethical Dilemma situations.
- Analyzing model behavior when presented with Unusual Access patterns.
Strengths
- Contains 106 million tokens, providing substantial volume for evaluation.
- Documents are generated using state-of-the-art models (GPT-4.1 and GPT-5) for specific traits.
- Covers seven distinct evaluation traits, including Verifiable Structure and Ethical Dilemma.
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
- compass-group-tue
- Collection Method
- Documents were generated using the pipeline of Wang et al. (2025) with GPT-4.1 and GPT-5.
- Freshness
- Last updated 2026-05-26 07:53:46; freshness should be verified.