Dataset Summary

This dataset is the RL training dataset for GoLongRL, targeting long-context capabilities of language models. It contains 23K training samples in total, with 9 types of reward functions.

You can load the dataset using:

from datasets import load_dataset
dataset = load_dataset("Kwai-Klear/GoLongRL")

See our paper and GitHub repository for more details.

📌 Overview

GoLongRL-4B achieves strong long-context performance at the 4B scale.

GoLongRL-30B-A3B achieves strong long-context performance at the 30B scale.

Data Fields

• data_source (string) - The identifier of the data source.
• prompt (list of dict) - The input prompt, stored as a list of message objects in chat format.
• ability (string) - Corresponds to the 9 task categories / reward types described in the paper.
• reward_model (dict) — Information about the ground truth or reward signal.
  • ground_truth (dict): The gold label used for RL training, containing three fields: doc_ids, golden_label, and summary. The summary field indicates the language of the training sample. Different task types require different data formats, utilizing either the doc_ids or golden_label field accordingly.
    • doc_ids (list): The gold label in list format, used for reward computation metrics such as F1 and NDCG.
    • golden_label (string): The gold label in string format, used for reward computation in tasks such as multiple-choice and fill-in-the-blank questions.
    • summary (string): Indicates the language of the sample. Only two possible values: "Chinese" and "English".
• extra_info (dict) - Information about the ground truth, question, and reward mode.
  • ground_truth (dict): Identical to the ground_truth field within reward_model.
  • question (string): The question portion of the training sample (excluding the document context).

Citation

If you find this work helpful, please cite our paper:

@misc{lv2026golongrlcapabilityorientedlongcontext,
      title={GoLongRL: Capability-Oriented Long Context Reinforcement Learning with Multitask Alignment}, 
      author={Minxuan Lv and Tiehua Mei and Tanlong Du and Junmin Chen and Zhenpeng Su and Ziyang Chen and Ziqi Wang and Zhennan Wu and Ruotong Pan and jian Liang and Ruiming Tang and Han Li},
      year={2026},
      eprint={2605.19577},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.19577}, 
}