A synthetic benchmark dataset for evaluating hallucination detection models on long documents (8K-24K tokens). This dataset is specifically designed to test models that can handle contexts beyond the typical 8K token limit.
Dataset Summary
Property
Value
Total samples
3,366
Token range
8,005 - 23,998
Average tokens
17,852
Hallucinated
1,681 (49.9%)
Supported
1,685 (50.1%)
Splits… See the full description on the dataset page: https://huggingface.co/datasets/llm-semantic-router/longcontext-haldetect.
