Generalized Hallucination Risk Level 100K Prompts

108,000 synthetically generated prompts are labeled for hallucination risk (LOW/MEDIUM/HIGH) via a multi-model consensus protocol. The dataset was created by author hb0810 to train pre-inference classifiers for Large Language Models. It was last updated on May 20, 2026.

Use Cases

• Train classifiers to predict hallucination risk based on prompt content.
• Benchmark the performance of different LLMs in evaluating prompt difficulty.
• Study the relationship between prompt topics and hallucination risk levels.
• Develop pre-inference filters to route high-risk prompts to more robust models.
• Analyze the consensus and disagreement patterns among multiple LLM evaluators.

Strengths

• Contains approximately 108,000 labeled prompts, providing substantial scale for model training.
• Labels were assigned via a protocol using three independent LLMs to reach consensus.
• Prompts were synthetically generated to cover a wide range of topics and difficulty levels.

Limitations

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

Provenance

Source

hb0810 on Hugging Face

Collection Method

Prompts were synthetically generated and labeled via a multi-model consensus protocol.

Freshness

Last updated 2026-05-20 11:16:29; freshness should be verified.