ChronoInstruct-SFT is the first chronologically consistent instruction-following dataset designed to remove lookahead bias in large language models.Each prompt–response pair is verified to contain only knowledge available before the year 2000, ensuring that no future information leaks into training.
The dataset was constructed as part of the paper:
Songrun He, Linying Lv, Asaf Manela, Jimmy Wu (2025). “Chronologically Consistent Generative AI.”Washington University… See the full description on the dataset page:
https://huggingface.co/datasets/manelalab/ChronoInstruct-SFT-v1.
