L-CITEEVAL: DO LONG-CONTEXT MODELS TRULY LEVERAGE CONTEXT FOR RESPONDING?

Paper   Github   Zhihu

Benchmark Quickview

L-CiteEval is a multi-task long-context understanding with citation benchmark, covering 5 task categories, including single-document question answering, multi-document question answering, summarization, dialogue understanding, and synthetic tasks, encompassing 11 different long-context tasks. The context lengths for these tasks range from 8K to 48K.

Data Prepare

Load Data

from datasets import load_dataset

datasets = ["narrativeqa", "natural_questions", "hotpotqa", "2wikimultihopqa", "goc_report", "multi_news", "qmsum", "locomo", "dialsim", "counting_stars", "niah"]

for dataset in datasets:
    ### Load L-CiteEval
    data = load_dataset('Jonaszky123/L-CiteEval', f"L-CiteEval-Data_{dataset}")

    ### Load L-CiteEval-Length
    data = load_dataset('Jonaszky123/L-CiteEval', f"L-CiteEval-Length_{dataset}")

    ### Load L-CiteEval-Hardness
    data = load_dataset('Jonaszky123/L-CiteEval', f"L-CiteEval-Hardness_{dataset}")

All data in L-CiteEval follows the format below:

{
    "id": "The identifier for the data entry",
    "question": "The task question, such as for single-document QA. In summarization tasks, this may be omitted",
    "answer": "The correct or expected answer to the question, used for evaluating correctness",
    "docs": "Context divided into fixed-length chunks"
    "length": "The context length"
    "hardness": "The level of difficulty in L-CiteEval-Hardness, which can be easy, medium and hard"
}

You can find evaluation code in our github.

Citation

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

@misc{tang2024lciteeval,
    title={L-CiteEval: Do Long-Context Models Truly Leverage Context for Responding?},
    author={Zecheng Tang and Keyan Zhou and Juntao Li and Baibei Ji and Jianye Hou and Min Zhang},
    year={2024},
    eprint={2410.02115},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}