Hugging Face's logo

Datasets:
zai-org
/
LongBench

Tasks:
Question Answering
Text Generation
Summarization
Languages:
English
Chinese
Size:
1K<n<10K
ArXiv:
Tags:
Long Context
Dataset card Files
xet
Community
6
nielsr HF Staff commited on
Commit
848a1b7
·
verified ·
1 Parent(s): 5e628be

Improve: Add related work (AttnTrace) and library_name

Browse files

This PR improves the `THUDM/LongBench` dataset card by:
- Adding `library_name: datasets` to the metadata for better discoverability and standard Hugging Face Hub practices.
- Introducing a "Used In / Related Work" section that links to the [AttnTrace paper](https://huggingface.co/papers/2508.03793) and its [Hugging Face Spaces demo](https://huggingface.co/spaces/SecureLLMSys/AttnTrace). This highlights an important application of the `LongBench` dataset in recent research on context traceback for LLMs.
- Updating the Arxiv link for the original `LongBench` paper from `.pdf` to `.abs` for a canonical URL.

Files changed (1) hide show
  1. README.md +70 -63
README.md CHANGED
@@ -1,16 +1,18 @@
1
  ---
 
 
 
 
 
2
  task_categories:
3
  - question-answering
4
  - text-generation
5
  - summarization
6
  - text-classification
7
- language:
8
- - en
9
- - zh
10
  tags:
11
  - Long Context
12
- size_categories:
13
- - 1K<n<10K
14
  ---
15
 
16
  # Introduction
@@ -22,7 +24,7 @@ We are fully aware of the potentially high costs involved in the model evaluatio
22
  LongBench includes 14 English tasks, 5 Chinese tasks, and 2 code tasks, with the average length of most tasks ranging from 5k to 15k, and a total of 4,750 test data. For detailed statistics and construction methods of LongBench tasks, please refer [here](task.md). In addition, we provide LongBench-E, a test set with a more uniform length distribution constructed by uniform sampling, with comparable amounts of data in the 0-4k, 4k-8k, and 8k+ length intervals to provide an analysis of the model's performance variations at different input lengths.
23
 
24
  Github Repo for LongBench: https://github.com/THUDM/LongBench
25
- Arxiv Paper for LongBench: https://arxiv.org/pdf/2308.14508.pdf
26
 
27
  # How to use it?
28
 
@@ -73,57 +75,57 @@ This repository provides data download for LongBench. If you wish to use this da
73
 
74
  # Task statistics
75
 
76
- | Task | Task Type | Eval metric | Avg len |Language | \#Sample |
77
  | :-------- | :-----------:| :-----------: |:-------: | :-----------: |:--------: |
78
- | HotpotQA | Multi-doc QA | F1 |9,151 |EN |200 |
79
- | 2WikiMultihopQA| Multi-doc QA | F1 |4,887 |EN |200 |
80
- | MuSiQue| Multi-doc QA | F1 |11,214 |EN |200 |
81
- | DuReader| Multi-doc QA | Rouge-L |15,768 |ZH |200 |
82
- | MultiFieldQA-en| Single-doc QA | F1 |4,559 |EN |150 |
83
- | MultiFieldQA-zh| Single-doc QA | F1 |6,701 |ZH |200 |
84
- | NarrativeQA| Single-doc QA | F1 |18,409 |EN |200 |
85
- | Qasper| Single-doc QA | F1 |3,619 |EN |200 |
86
- | GovReport| Summarization | Rouge-L |8,734 |EN |200 |
87
- | QMSum| Summarization | Rouge-L |10,614 |EN |200 |
88
- | MultiNews| Summarization | Rouge-L |2,113 |EN |200 |
89
- | VCSUM| Summarization | Rouge-L |15,380 |ZH |200 |
90
- | TriviaQA| Few shot | F1 |8,209 |EN |200 |
91
- | SAMSum| Few shot | Rouge-L |6,258 |EN |200 |
92
- | TREC| Few shot | Accuracy |5,177 |EN |200 |
93
- | LSHT| Few shot | Accuracy |22,337 |ZH |200 |
94
- | PassageRetrieval-en| Synthetic | Accuracy |9,289 |EN |200 |
95
- | PassageCount| Synthetic | Accuracy |11,141 |EN |200 |
96
- | PassageRetrieval-zh | Synthetic | Accuracy |6,745 |ZH |200 |
97
- | LCC| Code | Edit Sim |1,235 |Python/C#/Java |500 |
98
- | RepoBench-P| Code | Edit Sim |4,206 |Python/Java |500 |
99
 
100
  > Note: In order to avoid discrepancies caused by different tokenizers, we use the word count (using Python's split function) to calculate the average length of English datasets and code datasets, and use the character count to calculate the average length of Chinese datasets.
101
 
102
  # Task description
103
- | Task | Task Description |
104
  | :---------------- | :----------------------------------------------------------- |
105
- | HotpotQA | Answer related questions based on multiple given documents |
106
- | 2WikiMultihopQA | Answer related questions based on multiple given documents |
107
- | MuSiQue | Answer related questions based on multiple given documents |
108
- | DuReader | Answer related Chinese questions based on multiple retrieved documents |
109
- | MultiFieldQA-en | Answer English questions based on a long article, which comes from a relatively diverse field |
110
- | MultiFieldQA-zh | Answer Chinese questions based on a long article, which comes from a relatively diverse field |
111
- | NarrativeQA | Answer questions based on stories or scripts, including understanding of important elements such as characters, plots, themes, etc. |
112
- | Qasper | Answer questions based on a NLP research paper, questions proposed and answered by NLP practitioners |
113
- | GovReport | A summarization task that requires summarizing government work reports |
114
- | MultiNews | A multi-doc summarization that requires summarizing over multiple news |
115
- | QMSum | A summarization task that requires summarizing meeting records based on user queries |
116
- | VCSUM | A summarization task that requires summarizing Chinese meeting records |
117
- | SAMSum | A dialogue summarization task, providing several few-shot examples |
118
- | TriviaQA | Single document question answering task, providing several few-shot examples |
119
- | NQ | Single document question answering task, providing several few-shot examples |
120
- | TREC | A classification task that requires categorizing questions, includes 50 categories in total |
121
- | LSHT | A Chinese classification task that requires categorizing news, includes 24 categories in total |
122
  | PassageRetrieval-en | Given 30 English Wikipedia paragraphs, determine which paragraph the given summary corresponds to |
123
  | PassageCount | Determine the total number of different paragraphs in a given repetitive article |
124
  | PassageRetrieval-zh | Given several Chinese paragraphs from the C4 data set, determine which paragraph the given abstract corresponds to |
125
- | LCC | Given a long piece of code, predict the next line of code |
126
- | RepoBench-P | Given code in multiple files within a GitHub repository (including cross-file dependencies), predict the next line of code |
127
 
128
  # Task construction
129
  > Note: For all tasks constructed from existing datasets, we use data from the validation or test set of the existing dataset (except for VCSUM).
@@ -140,24 +142,29 @@ This repository provides data download for LongBench. If you wish to use this da
140
  - For the [LCC](https://arxiv.org/abs/2306.14893) task, we sample from the original code completion dataset. In the [RepoBench-P](https://arxiv.org/abs/2306.03091) task, we select the most challenging XF-F (Cross-File-First) setting from the original dataset and refer to the Oracle-Filled scenario in the paper. For each original piece of data, we randomly extract multiple cross-file code snippets, including the gold cross-file code snippet, and concatenate them as input, requiring the model to effectively use cross-file code for completion.
141
 
142
  # LongBench-E statistics
143
- | Task | Task Type | \#data in 0-4k | \#data in 4-8k | \#data in 8k+|
144
  | :--------- | :-----------:| :-----------: |:---------: | :-------------: |
145
- | HotpotQA | Multi-doc QA | 100 |100 |100 |
146
- | 2WikiMultihopQA| Multi-doc QA | 100 |100 |100 |
147
- | MultiFieldQA-en| Single-doc QA | 67 |70 |13 |
148
- | Qasper| Single-doc QA | 100 |100 |24 |
149
- | GovReport| Summarization | 100 |100 |100 |
150
- | MultiNews| Summarization | 100 |100 |94 |
151
- | TriviaQA| Few shot | 100 |100 |100 |
152
- | SAMSum| Few shot | 100 |100 |100 |
153
- | TREC| Few shot | 100 |100 |100 |
154
- | PassageRetrieval-en| Synthetic | 100 |100 |100 |
155
- | PassageCount| Synthetic | 100 |100 |100 |
156
- | LCC| Code | 100 |100 |100 |
157
- | RepoBench-P| Code | 100 |100 |100 |
 
 
 
 
 
158
 
159
  # Citation
160
- ```
161
  @misc{bai2023longbench,
162
  title={LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding},
163
  author={Yushi Bai and Xin Lv and Jiajie Zhang and Hongchang Lyu and Jiankai Tang and Zhidian Huang and Zhengxiao Du and Xiao Liu and Aohan Zeng and Lei Hou and Yuxiao Dong and Jie Tang and Juanzi Li},
 
1
  ---
2
+ language:
3
+ - en
4
+ - zh
5
+ size_categories:
6
+ - 1K<n<10K
7
  task_categories:
8
  - question-answering
9
  - text-generation
10
  - summarization
11
  - text-classification
 
 
 
12
  tags:
13
  - Long Context
14
+ library_name:
15
+ - datasets
16
  ---
17
 
18
  # Introduction
 
24
  LongBench includes 14 English tasks, 5 Chinese tasks, and 2 code tasks, with the average length of most tasks ranging from 5k to 15k, and a total of 4,750 test data. For detailed statistics and construction methods of LongBench tasks, please refer [here](task.md). In addition, we provide LongBench-E, a test set with a more uniform length distribution constructed by uniform sampling, with comparable amounts of data in the 0-4k, 4k-8k, and 8k+ length intervals to provide an analysis of the model's performance variations at different input lengths.
25
 
26
  Github Repo for LongBench: https://github.com/THUDM/LongBench
27
+ Arxiv Paper for LongBench: [https://arxiv.org/abs/2308.14508](https://arxiv.org/abs/2308.14508)
28
 
29
  # How to use it?
30
 
 
75
 
76
  # Task statistics
77
 
78
+ | Task | Task Type | Eval metric | Avg len |Language | #Sample |
79
  | :-------- | :-----------:| :-----------: |:-------: | :-----------: |:--------: |
80
+ | HotpotQA | Multi-doc QA | F1 |9,151 |EN |200 |
81
+ | 2WikiMultihopQA| Multi-doc QA | F1 |4,887 |EN |200 |
82
+ | MuSiQue| Multi-doc QA | F1 |11,214 |EN |200 |
83
+ | DuReader| Multi-doc QA | Rouge-L |15,768 |ZH |200 |
84
+ | MultiFieldQA-en| Single-doc QA | F1 |4,559 |EN |150 |
85
+ | MultiFieldQA-zh| Single-doc QA | F1 |6,701 |ZH |200 |
86
+ | NarrativeQA| Single-doc QA | F1 |18,409 |EN |200 |
87
+ | Qasper| Single-doc QA | F1 |3,619 |EN |200 |
88
+ | GovReport| Summarization | Rouge-L |8,734 |EN |200 |
89
+ | QMSum| Summarization | Rouge-L |10,614 |EN |200 |
90
+ | MultiNews| Summarization | Rouge-L |2,113 |EN |200 |
91
+ | VCSUM| Summarization | Rouge-L |15,380 |ZH |200 |
92
+ | TriviaQA| Few shot | F1 |8,209 |EN |200 |
93
+ | SAMSum| Few shot | Rouge-L |6,258 |EN |200 |
94
+ | TREC| Few shot | Accuracy |5,177 |EN |200 |
95
+ | LSHT| Few shot | Accuracy |22,337 |ZH |200 |
96
+ | PassageRetrieval-en| Synthetic | Accuracy |9,289 |EN |200 |
97
+ | PassageCount| Synthetic | Accuracy |11,141 |EN |200 |
98
+ | PassageRetrieval-zh | Synthetic | Accuracy |6,745 |ZH |200 |
99
+ | LCC| Code | Edit Sim |1,235 |Python/C#/Java |500 |
100
+ | RepoBench-P| Code | Edit Sim |4,206 |Python/Java |500 |
101
 
102
  > Note: In order to avoid discrepancies caused by different tokenizers, we use the word count (using Python's split function) to calculate the average length of English datasets and code datasets, and use the character count to calculate the average length of Chinese datasets.
103
 
104
  # Task description
105
+ | Task | Task Description |
106
  | :---------------- | :----------------------------------------------------------- |
107
+ | HotpotQA | Answer related questions based on multiple given documents |
108
+ | 2WikiMultihopQA | Answer related questions based on multiple given documents |
109
+ | MuSiQue | Answer related questions based on multiple given documents |
110
+ | DuReader | Answer related Chinese questions based on multiple retrieved documents |
111
+ | MultiFieldQA-en | Answer English questions based on a long article, which comes from a relatively diverse field |
112
+ | MultiFieldQA-zh | Answer Chinese questions based on a long article, which comes from a relatively diverse field |
113
+ | NarrativeQA | Answer questions based on stories or scripts, including understanding of important elements such as characters, plots, themes, etc. |
114
+ | Qasper | Answer questions based on a NLP research paper, questions proposed and answered by NLP practitioners |
115
+ | GovReport | A summarization task that requires summarizing government work reports |
116
+ | MultiNews | A multi-doc summarization that requires summarizing over multiple news |
117
+ | QMSum | A summarization task that requires summarizing meeting records based on user queries |
118
+ | VCSUM | A summarization task that requires summarizing Chinese meeting records |
119
+ | SAMSum | A dialogue summarization task, providing several few-shot examples |
120
+ | TriviaQA | Single document question answering task, providing several few-shot examples |
121
+ | NQ | Single document question answering task, providing several few-shot examples |
122
+ | TREC | A classification task that requires categorizing questions, includes 50 categories in total |
123
+ | LSHT | A Chinese classification task that requires categorizing news, includes 24 categories in total |
124
  | PassageRetrieval-en | Given 30 English Wikipedia paragraphs, determine which paragraph the given summary corresponds to |
125
  | PassageCount | Determine the total number of different paragraphs in a given repetitive article |
126
  | PassageRetrieval-zh | Given several Chinese paragraphs from the C4 data set, determine which paragraph the given abstract corresponds to |
127
+ | LCC | Given a long piece of code, predict the next line of code |
128
+ | RepoBench-P | Given code in multiple files within a GitHub repository (including cross-file dependencies), predict the next line of code |
129
 
130
  # Task construction
131
  > Note: For all tasks constructed from existing datasets, we use data from the validation or test set of the existing dataset (except for VCSUM).
 
142
  - For the [LCC](https://arxiv.org/abs/2306.14893) task, we sample from the original code completion dataset. In the [RepoBench-P](https://arxiv.org/abs/2306.03091) task, we select the most challenging XF-F (Cross-File-First) setting from the original dataset and refer to the Oracle-Filled scenario in the paper. For each original piece of data, we randomly extract multiple cross-file code snippets, including the gold cross-file code snippet, and concatenate them as input, requiring the model to effectively use cross-file code for completion.
143
 
144
  # LongBench-E statistics
145
+ | Task | Task Type | #data in 0-4k | #data in 4-8k | #data in 8k+|
146
  | :--------- | :-----------:| :-----------: |:---------: | :-------------: |
147
+ | HotpotQA | Multi-doc QA | 100 |100 |100 |
148
+ | 2WikiMultihopQA| Multi-doc QA | 100 |100 |100 |
149
+ | MultiFieldQA-en| Single-doc QA | 67 |70 |13 |
150
+ | Qasper| Single-doc QA | 100 |100 |24 |
151
+ | GovReport| Summarization | 100 |100 |100 |
152
+ | MultiNews| Summarization | 100 |100 |94 |
153
+ | TriviaQA| Few shot | 100 |100 |100 |
154
+ | SAMSum| Few shot | 100 |100 |100 |
155
+ | TREC| Few shot | 100 |100 |100 |
156
+ | PassageRetrieval-en| Synthetic | 100 |100 |100 |
157
+ | PassageCount| Synthetic | 100 |100 |100 |
158
+ | LCC| Code | 100 |100 |100 |
159
+ | RepoBench-P| Code | 100 |100 |100 |
160
+
161
+ # Used In / Related Work
162
+ This dataset has been utilized in recent research, such as:
163
+ - [**AttnTrace: Attention-based Context Traceback for Long-Context LLMs**](https://huggingface.co/papers/2508.03793)
164
+ - Project page / Demo: [https://huggingface.co/spaces/SecureLLMSys/AttnTrace](https://huggingface.co/spaces/SecureLLMSys/AttnTrace)
165
 
166
  # Citation
167
+ ```bib
168
  @misc{bai2023longbench,
169
  title={LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding},
170
  author={Yushi Bai and Xin Lv and Jiajie Zhang and Hongchang Lyu and Jiankai Tang and Zhidian Huang and Zhengxiao Du and Xiao Liu and Aohan Zeng and Lei Hou and Yuxiao Dong and Jie Tang and Juanzi Li},