LongBench

Dataset Summary

LongBench is a bilingual, multitask benchmark for evaluating long-context understanding in large language models. It covers long-text application scenarios including single-document question answering, multi-document question answering, summarization, few-shot learning, synthetic long-context tasks, and code completion.

This Hugging Face dataset repository repackages locally downloaded LongBench JSONL files into a clean, typed, data-only Hugging Face dataset layout with one configuration per task. The goal of this repackaging is ease of use, reproducibility, dataset viewer compatibility, efficient loading, and convenient downstream evaluation. The dataset content, task design, and original benchmark are attributed to the LongBench authors and the THUDM LongBench project.

This repository keeps the LongBench records in their original task-level shape while publishing them as typed Parquet configs. The split is always test; the configuration name is the LongBench task name. That means the benchmark can be loaded with plain load_dataset(...) calls, without the legacy dataset script or a manual data.zip step.

The language metadata above is intentionally limited to Hub-valid values. Row-level LongBench labels are preserved in the data and include natural languages (en, zh) plus code labels (python, java, csharp) for the completion tasks.

This repackaging also adds token counts to every row: input_tokens, context_tokens, and total_tokens, computed with cl100k_base. LongBench is specifically about long-context behavior, so the added counts make it easier to filter by real prompt size, inspect outliers, and compare tasks without guessing from the original mixed length field.

Original Source and Attribution

• Original project: https://github.com/THUDM/LongBench
• Paper: https://arxiv.org/abs/2308.14508
• Original authors: Yushi Bai, Xin Lv, Jiajie Zhang, Hongchang Lyu, Jiankai Tang, Zhidian Huang, Zhengxiao Du, Xiao Liu, Aohan Zeng, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li.
• This repository: GinkgoQ/LongBench
• Packaging contribution: typed Parquet configs, Hub-valid metadata, local build metadata, and per-row cl100k_base token counts.
• Attribution note: the benchmark design and source records remain LongBench.

Dataset Structure

Each LongBench task is exposed as a separate Hugging Face configuration. Every configuration has a single test split.

Schema

Field	Type	Description
input	string	The model-facing question, prompt, instruction, query, or task input.
context	string	The long-context document, passage set, dialogue, code context, or retrieved evidence.
answers	list[string]	Gold reference answer or answer candidates.
length	int32	Source-provided context length metadata.
input_tokens	int32	Number of cl100k_base tokens in input. Added by this repackaging.
context_tokens	int32	Number of cl100k_base tokens in context. Added by this repackaging.
total_tokens	int32	Number of cl100k_base tokens in input plus context, counted with one newline separator when input is non-empty. Added by this repackaging.
dataset	string	Original dataset/task label from LongBench.
language	string	Source-provided language label.
all_classes	list[string]	Candidate class labels when applicable. Empty for tasks where this is not used.
_id	string	Original example identifier, preserved when available.

Configurations

Configuration	Task Group	Examples	Languages	Mean Source Length	Mean Tokens	Max Tokens
narrativeqa	Single-Document QA	200	en	18404.94	29790.89	65293
qasper	Single-Document QA	200	en	3618.7	4932.52	21129
multifieldqa_en	Single-Document QA	150	en	4558.7	6951.12	14962
multifieldqa_zh	Single-Document QA	200	zh	6700.68	7296.32	14962
hotpotqa	Multi-Document QA	200	en	9149.22	12812.86	16346
2wikimqa	Multi-Document QA	200	en	4885.31	7133.24	16356
musique	Multi-Document QA	200	en	11017.66	15595.57	16353
dureader	Multi-Document QA	200	zh	15768.04	17605.0	32255
gov_report	Summarization	200	en	8169.36	10242.25	51394
qmsum	Summarization	200	en	10545.94	13868.7	30389
multi_news	Summarization	200	en	2113.49	2609.06	13935
vcsum	Summarization	200	zh	15147.02	16896.71	49027
trec	Few-Shot Learning	200	en	5176.36	6768.22	11382
triviaqa	Few-Shot Learning	200	en	8209.3	11771.0	23349
samsum	Few-Shot Learning	200	en	6258.35	9155.56	17981
lsht	Few-Shot Learning	200	zh	22332.62	26322.06	51727
passage_count	Synthetic Tasks	200	en	11140.59	14898.67	28965
passage_retrieval_en	Synthetic Tasks	200	en	9287.97	12471.94	15188
passage_retrieval_zh	Synthetic Tasks	200	zh	6745.15	7765.06	10736
lcc	Code Completion	500	csharp, java, python	1235.28	3165.98	30150
repobench-p	Code Completion	500	java, python	4205.93	10813.41	39128
qasper_e	LongBench-E	224	en	4620.48	6218.5	21129
multifieldqa_en_e	LongBench-E	150	en	4558.7	6951.12	14962
hotpotqa_e	LongBench-E	300	en	6657.96	9470.88	16329
2wikimqa_e	LongBench-E	300	en	6146.54	8874.2	16333
gov_report_e	LongBench-E	300	en	7140.79	8160.53	27686
multi_news_e	LongBench-E	294	en	5999.31	7883.37	38322
trec_e	LongBench-E	300	en	6259.26	8181.84	17185
triviaqa_e	LongBench-E	300	en	6684.6	9693.12	36228
samsum_e	LongBench-E	300	en	6170.48	9035.07	18223
passage_count_e	LongBench-E	300	en	6117.3	8232.71	22952
passage_retrieval_en_e	LongBench-E	300	en	6115.38	8185.44	14490
lcc_e	LongBench-E	300	csharp, java, python	5546.3	13516.84	49200
repobench-p_e	LongBench-E	300	java, python	6067.31	15312.48	41008

Task Groups

Single-Document QA

• Configurations: narrativeqa, qasper, multifieldqa_en, multifieldqa_zh
• Examples: 750

Multi-Document QA

• Configurations: hotpotqa, 2wikimqa, musique, dureader
• Examples: 800

Summarization

• Configurations: gov_report, qmsum, multi_news, vcsum
• Examples: 800

Few-Shot Learning

• Configurations: trec, triviaqa, samsum, lsht
• Examples: 800

Synthetic Tasks

• Configurations: passage_count, passage_retrieval_en, passage_retrieval_zh
• Examples: 600

Code Completion

• Configurations: lcc, repobench-p
• Examples: 1000

LongBench-E

• Configurations: qasper_e, multifieldqa_en_e, hotpotqa_e, 2wikimqa_e, gov_report_e, multi_news_e, trec_e, triviaqa_e, samsum_e, passage_count_e, passage_retrieval_en_e, lcc_e, repobench-p_e
• Examples: 3668

Languages

Row-level labels: csharp, en, java, python, zh

Hub metadata labels: code, en, zh

Code labels preserved in rows: csharp, java, python

Source Dataset Labels

2wikimqa, 2wikimqa_e, dureader, gov_report, gov_report_e, hotpotqa, hotpotqa_e, lcc, lcc_e, lsht, multi_news, multi_news_e, multifieldqa_en, multifieldqa_en_e, multifieldqa_zh, musique, narrativeqa, passage_count, passage_count_e, passage_retrieval_en, passage_retrieval_en_e, passage_retrieval_zh, qasper, qasper_e, qmsum, repobench-p, repobench-p_e, samsum, samsum_e, trec, trec_e, triviaqa, triviaqa_e, vcsum

Token Counts

Token counts are generated during packaging with cl100k_base:

• input_tokens: tokens in the task input or question.
• context_tokens: tokens in the long context.
• total_tokens: tokens in the combined input/context prompt.

Across this build, the mean total_tokens is 10450.43 and the largest row has 65293 tokens.

Loading

Load one task:

from datasets import load_dataset

dataset = load_dataset("GinkgoQ/LongBench", "narrativeqa", split="test")
print(dataset)
print(dataset[0])

Load multiple tasks:

from datasets import load_dataset

tasks = [
  "narrativeqa",
  "qasper",
  "multifieldqa_en",
  "multifieldqa_zh",
  "hotpotqa",
  "2wikimqa",
  "musique",
  "dureader",
  "gov_report",
  "qmsum",
  "multi_news",
  "vcsum",
  "trec",
  "triviaqa",
  "samsum",
  "lsht",
  "passage_count",
  "passage_retrieval_en",
  "passage_retrieval_zh",
  "lcc",
  "repobench-p",
  "qasper_e",
  "multifieldqa_en_e",
  "hotpotqa_e",
  "2wikimqa_e",
  "gov_report_e",
  "multi_news_e",
  "trec_e",
  "triviaqa_e",
  "samsum_e",
  "passage_count_e",
  "passage_retrieval_en_e",
  "lcc_e",
  "repobench-p_e"
]

datasets_by_task = {
    task: load_dataset("GinkgoQ/LongBench", task, split="test")
    for task in tasks
}

Load all available configurations dynamically:

from datasets import get_dataset_config_names, load_dataset

repo_id = "GinkgoQ/LongBench"
configs = get_dataset_config_names(repo_id)

datasets_by_task = {
    config: load_dataset(repo_id, config, split="test")
    for config in configs
}

Example Record

{
  "input": "...",
  "context": "...",
  "answers": ["..."],
  "length": 12345,
  "input_tokens": 12,
  "context_tokens": 6789,
  "total_tokens": 6802,
  "dataset": "narrativeqa",
  "language": "en",
  "all_classes": [],
  "_id": "..."
}

Intended Use

This dataset is intended for:

• Long-context language model evaluation
• Benchmarking retrieval-augmented and long-context systems
• Comparing performance across long-document QA, multi-document QA, summarization, classification, synthetic reasoning, and code-completion tasks
• Reproducible evaluation workflows using the Hugging Face datasets library

Out-of-Scope Use

This dataset should not be used as the sole evidence for claims about general model safety, factuality, robustness, legal compliance, medical reliability, or deployment readiness. It is an evaluation benchmark and should be combined with domain-specific tests when used for production model assessment.

Data Fields

input

The model-facing user query, prompt, question, task instruction, or completion prefix.

context

The long context provided to the model. Depending on the task, this may contain documents, passages, reports, dialogue, retrieved evidence, or source code.

answers

Reference answer list. Some tasks may include multiple valid answers.

length

Source-provided length metadata.

input_tokens

Number of cl100k_base tokens in input, added by this packaging script.

context_tokens

Number of cl100k_base tokens in context, added by this packaging script.

total_tokens

Number of cl100k_base tokens in the combined input/context prompt. When input is non-empty, the counter uses input + "\n" + context; otherwise it counts context.

dataset

Original dataset or task label.

language

Source-provided language metadata.

all_classes

Candidate labels for classification-style tasks. Empty when not applicable.

_id

Original example identifier when available. If an identifier was missing in a local source row, this build pipeline generated a deterministic fallback identifier using the task name and row index.

Build Details

This repository was generated automatically from local JSONL files using a validation and conversion pipeline.

• Build timestamp UTC: 2026-05-24T08:44:19.240990+00:00
• Source directory: /home/arman/project/LongBench/LongBench_data/data
• Number of configurations: 34
• Total examples: 8418
• File format: Parquet
• Split: test
• Schema: fixed typed schema shared by all configurations
• Validation mode: strict
• Max shard size: 500MB
• Token count method: cl100k_base

Processing Pipeline

The build pipeline performs the following steps:

1. Detects available LongBench JSONL files.
2. Validates task names against the known LongBench task list.
3. Reads each JSONL file line by line.
4. Validates JSON syntax and row object type.
5. Normalizes all original fields into a consistent Hugging Face schema.
6. Adds input_tokens, context_tokens, and total_tokens with cl100k_base.
7. Preserves the original LongBench fields.
8. Converts each task into a typed Hugging Face Dataset.
9. Writes each task as Parquet under data/<config>/test-*.parquet.
10. Generates this dataset card dynamically from the detected files and statistics.
11. Generates dataset_infos.json and build_metadata.json.
12. Optionally creates the Hugging Face dataset repository.
13. Uploads the generated repository folder to the Hugging Face Hub.
14. Optionally performs a remote smoke test with load_dataset.

Validation Notes

The build script supports strict and non-strict modes.

In strict mode, the script fails if required fields are missing, if input or context are empty, if length is negative, or if list-like fields cannot be normalized.

In non-strict mode, the script preserves maximum compatibility by filling missing optional values with deterministic defaults where possible.

Citation

If you use this repackaged dataset, cite the original LongBench paper:

@article{bai2023longbench,
  title={LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding},
  author={Bai, Yushi and Lv, Xin and Zhang, Jiajie and Lyu, Hongchang and Tang, Jiankai and Huang, Zhidian and Du, Zhengxiao and Liu, Xiao and Zeng, Aohan and Hou, Lei and Dong, Yuxiao and Tang, Jie and Li, Juanzi},
  journal={arXiv preprint arXiv:2308.14508},
  year={2023}
}

License

This repository uses the license metadata apache-2.0. Users should verify licensing and redistribution requirements against the original LongBench project and any upstream datasets included in LongBench before public redistribution or commercial usage.

Acknowledgements

All benchmark design, task construction, and source data attribution belong to the LongBench authors and the THUDM LongBench project. This repository only repackages the source files for easier loading and use through the Hugging Face Hub.