ChronoGPT-Instruct is a family of chronologically consistent, instruction-following large language models (LLMs) that eliminate lookahead bias by training exclusively on time-stamped data available before a fixed knowledge-cutoff date τ.
Each ChronoGPT-Instruct-τ extends the ChronoGPT-τ base models through supervised instruction fine-tuning while strictly maintaining temporal separation from all post-τ information.

These models provide the research community with a transparent, replicable benchmark for testing lookahead-bias-free prediction in economics, finance, and other time-sensitive domains.

Each model’s corpus satisfies chronologically consistency in both pretraining and instruction-finetuning phases. Texts dated after the model year are excluded, ensuring zero overlap with evaluation data. A GPT-4.1 classifier screens every instruction-response pair.

Only English, non-code entries with pre-2000 content (classifier label = 0 & confidence = 10) are retained.

We release the SFT dataset at https://huggingface.co/datasets/manelalab/ChronoInstruct-SFT.

You can try ChronoGPT-instruct directly in your browser via Google Colab:

@article{He_Lv_Manela_Wu_chronogpt_2025,
  title={Instruction Tuning Chronologically Consistent Language Models},
  author={He, Songrun and Lv, Linying and Manela, Asaf and Wu, Jimmy},
  journal={Working Paper},
  year={2025}
}