Awesome

InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens

Introduction

Welcome to InfiniteBench, a cutting-edge benchmark tailored for evaluating the capabilities of language models to process, understand, and reason over super long contexts (100k+ tokens). Long contexts are crucial for enhancing applications with LLMs and achieving high-level interaction. InfiniteBench is designed to push the boundaries of language models by testing them against a context length of 100k+, which is 10 times longer than traditional datasets.

Features

• Loooong Context: InfiniteBench is a pioneer in testing language models with a context length of 100k+, offering an unparalleled challenge in the field.
• Diverse Domain: The benchmark comprises 12 unique tasks, each crafted to assess different aspects of language processing and comprehension in extended contexts.
• Specialized Test: InfiniteBench consists of tasks that state-of-the-art LLMs are known to be capable of when using shorter context. This ensures that the performance degradation is only caused by the length of the contexts.
• Real-World and Synthetic Scenarios: The tasks are a mix of real-world scenarios and synthetic constructs, ensuring a comprehensive evaluation of models. Real-world scenarios make the test pragmatic, and synthetic ones leave the space for extending the context length further with ease.

Task Composition

How to Download Data

Click here to download data from 🤗 Huggingface directly: https://huggingface.co/datasets/xinrongzhang2022/InfiniteBench

Using 🤗 Datasets

Alternatively, you can download using the 🤗 Datasets library as follows.

from datasets import load_dataset, Value, Sequence
ft = Features({"id": Value("int64"), "context": Value("string"), "input": Value("string"), "answer": Sequence(Value("string")), "options": Sequence(Value("string"))})
dataset = load_dataset("xinrongzhang2022/InfiniteBench", features=ft)

Using Scripts

cd InfiniteBench
bash scripts/download_dataset.sh

This will directly dump the data to data.

Evaluation Result

We evaluate SOTA proprietary and open-source LLMs, the result is as follows.

Note:

1. The evaluation code for YaRN-Mistral-7B is implemented by ourselves, and please contact us or submit an issue if there are any problems.

2. Kimi-Chat, Claude 2, and GPT-4 are evaluated using the official API with default configuration.

3. For Math.Calc, the values in the parentheses have a measurement unit of 0.01%. This is because it is easy to get a very low score on this task.

4. The metric for task Math.Find, Math.Calc, Code.Run, Code.Debug, En.Dia, En.MC, Retrieve.KV, Retrieve.Number, and Retrieve.PassKey is accuracy;

   The metric for task Zh.QA and En.QA are ROUGE F1 score;

   The metric for En.Sum is the rougeLsum score from the 🤗 Evaluate library.

Installation

pip install -r requirements.txt

How to Run

Download the dataset the data folder (or set the --data_dir argument to the location of the dataset). The data folder structure should be as follows.

InfiniteBench
├── data
│   ├── code_debug.jsonl
│   ├── code_run.jsonl
│   ├── kv_retrieval.jsonl
│   ├── longbook_choice_eng.jsonl
│   ├── longbook_qa_chn.jsonl
│   ├── longbook_qa_eng.jsonl
│   ├── longbook_sum_eng.jsonl
│   ├── longdialogue_qa_eng.jsonl
│   ├── math_calc.jsonl
│   ├── math_find.jsonl
│   ├── number_string.jsonl
│   ├── passkey.jsonl
│   └── construct_synthetic_dataset.py
...

Then, in the src folder, execute:

python eval_yarn_mistral.py --task kv_retrieval
python eval_gpt4.py --task longbook_sum_qa
python eval_rwkv.py --task passkey

The available tasks are:

Citation

This will be updated when our preprint paper is released.

@inproceedings{zhang-etal-2024-bench,
    title = "$\infty${B}ench: Extending Long Context Evaluation Beyond 100{K} Tokens",
    author = "Zhang, Xinrong  and
      Chen, Yingfa  and
      Hu, Shengding  and
      Xu, Zihang  and
      Chen, Junhao  and
      Hao, Moo  and
      Han, Xu  and
      Thai, Zhen  and
      Wang, Shuo  and
      Liu, Zhiyuan  and
      Sun, Maosong",
    editor = "Ku, Lun-Wei  and
      Martins, Andre  and
      Srikumar, Vivek",
    booktitle = "Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = aug,
    year = "2024",
    address = "Bangkok, Thailand",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.acl-long.814",
    pages = "15262--15277",
    abstract = "Processing and reasoning over long contexts is crucial for many practical applications of Large Language Models (LLMs), such as document comprehension and agent construction. Despite recent strides in making LLMs process contexts with more than 100K tokens, there is currently a lack of a standardized benchmark to evaluate this long-context capability. Existing public benchmarks typically focus on contexts around 10K tokens, limiting the assessment and comparison of LLMs in processing longer contexts. In this paper, we propose , the first LLM benchmark featuring an average data length surpassing 100K tokens. comprises synthetic and realistic tasks spanning diverse domains in English and Chinese. The tasks in are designed to require an understanding of long dependencies in contexts and make simply retrieving a limited number of passages from contexts not sufficient for these tasks. Based on , we evaluate several state-of-the-art LLMs tailored for processing long contexts. The experimental results indicate that existing long-context LLMs still require significant advancements to process 100K+ contexts effectively. Furthermore, we present three intriguing analyses regarding the behavior of LLMs processing long context. Our code and data is released.",
}

Acknowledgement

Thanks to Cong Feng, Zhongwu Zhai, Guoyang Zeng, Chenyang Song, Renjie Luo, Chaoqun He, Yuge Tu, Bowen Ping, Yujie Huang, Yudong Mei, Kaihuo Zhang, Weilin Zhao, Ao Sun, Yulin Chen, Ganqu Cui.

References