metadata
language:
- en
task_categories:
- text-generation
tags:
- long-context
prolong-data-64K
This dataset was used in the paper CoPE: Clipped RoPE as A Scalable Free Lunch for Long Context LLMs. It is based on the princeton-nlp/prolong-data-64K dataset.
- Paper: https://huggingface.co/papers/2602.05258
- GitHub Repository: https://github.com/hrlics/CoPE
Overview
CoPE (Clipped RoPE) is a plug-and-play enhancement of RoPE that softly clips unstable low-frequency components, delivering consistent gains both within the training context and during long-context extrapolation. This dataset was utilized for continued pre-training and supervised fine-tuning (SFT) to scale models (starting from Llama-3-8B) to a 64k context length.
Usage
According to the GitHub README, the data can be downloaded using:
git clone https://huggingface.co/datasets/haoranli-ml/prolong-data-64K datasets/long-context-65536
Citation
@article{li2026cope,
title={CoPE: Clipped RoPE as A Scalable Free Lunch for Long Context LLMs},
author={Li, Haoran and Ren, Sucheng and Yuille, Alan and Wang, Feng},
journal={arXiv preprint arXiv:2602.05258},
year={2026}
}