LLM Confidence Decoupling Benchmark Under Adversarial Pressure

347 questions across 3 adversarial variants form a benchmark for evaluating large language model confidence. It measures how model confidence changes under different types of adversarial pressure. The dataset's author and creation date are not specified.

Use Cases

• Benchmarking model confidence scores on the 347 base questions against adversarial variants.
• Analyzing confidence decoupling patterns across the 3 defined adversarial conditions.
• Training or fine-tuning models to maintain calibrated confidence under adversarial pressure using the provided question-answer pairs.

Strengths

• 347 distinct questions provide a foundation for evaluation.
• 3 adversarial variants enable analysis of confidence under specific pressure types.

Limitations

• The dataset size is limited to 347 core questions.
• Specifics on question content, answer formats, and adversarial construction methods are not provided.

Provenance

Source

Kaggle

Collection Method

Constructed for benchmarking LLM confidence, likely involving manual or automated generation of adversarial variants.