Datasets:

Modalities:
Text
Formats:
parquet
Languages:
English
ArXiv:
Libraries:
Datasets
pandas
License:
File size: 4,203 Bytes
974f7ec
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0dd7881
9df7e1c
fa6d628
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0dd7881
 
 
 
 
 
 
 
 
 
 
 
fa6d628
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
---
license: cc-by-sa-4.0
task_categories:
- text-generation
language:
- en
size_categories:
- 1K<n<10K
configs:
- config_name: github_prs
  data_files:
  - split: validation
    path: github_prs/validation-*
- config_name: numerical
  data_files:
  - split: validation
    path: numerical/validation-*
- config_name: poetry
  data_files:
  - split: validation
    path: poetry/validation-*
dataset_info:
- config_name: github_prs
  features:
  - name: id
    dtype: int64
  - name: original_context
    dtype: string
  - name: modified_context
    dtype: string
  - name: omitted_context
    sequence: string
  - name: omitted_index
    sequence: int64
  - name: metadata
    struct:
    - name: additions
      dtype: int64
    - name: author
      dtype: string
    - name: deletions
      dtype: int64
    - name: html_url
      dtype: string
    - name: issue_id
      dtype: int64
    - name: merged_at
      dtype: string
    - name: omission_probability
      dtype: float64
    - name: pr_number
      dtype: int64
    - name: repo
      dtype: string
    - name: title
      dtype: string
    - name: total_changes
      dtype: int64
  splits:
  - name: validation
    num_bytes: 9578209
    num_examples: 887
  download_size: 3431644
  dataset_size: 9578209
- config_name: numerical
  features:
  - name: id
    dtype: int64
  - name: original_context
    dtype: string
  - name: modified_context
    dtype: string
  - name: omitted_context
    sequence: string
  - name: omitted_index
    sequence: int64
  - name: metadata
    struct:
    - name: max_num
      dtype: int64
    - name: min_num
      dtype: int64
    - name: n_numbers
      dtype: int64
    - name: n_omitted
      dtype: int64
    - name: omission_probability
      dtype: float64
    - name: order
      dtype: string
    - name: step_size
      dtype: int64
  splits:
  - name: validation
    num_bytes: 7175806
    num_examples: 1200
  download_size: 4970832
  dataset_size: 7175806
- config_name: poetry
  features:
  - name: id
    dtype: int64
  - name: original_context
    dtype: string
  - name: modified_context
    dtype: string
  - name: omitted_context
    sequence: string
  - name: omitted_index
    sequence: int64
  - name: metadata
    struct:
    - name: n_omitted
      dtype: int64
    - name: omission_probability
      dtype: float64
    - name: poem_length
      dtype: int64
  splits:
  - name: validation
    num_bytes: 44217657
    num_examples: 1191
  download_size: 29473174
  dataset_size: 44217657
---
# AbsenceBench: Language Models Can't Tell What's Missing
[[paper]](https://arxiv.org/abs/2506.11440) [[code]](https://github.com/harvey-fin/absence-bench)

# Dataset Card for AbsenceBench

AbsenceBench covers three distinct domains:
- Poetry (realistic)
- Numerical sequences (synthetic)
- GitHub pull requests (realistic)

There are 4302 instances in total, with an average context length of 5K tokens.

## Data Structure
Each domain contains the following features:
- `original_context`: `string`
- `modified_context`: `string`
- `omitted_context`: `List[string]`
- `omitted_index`: `List[int]` 
- `metadata`: `Dict` Extra information regarding the dataset generation

### Data Split
Currently we provide only the `validation` split. Since all instances of AbsenceBench natually come up with labeled answers according to the task setting, we will not be able to provide an unlabeled `test` split.

## Dataset Construction
We provide the source data for generating the poetry domain, originally from the [Gutenberg Poetry Corpus](https://github.com/aparrish/gutenberg-poetry-corpus) in the `poetry_raw` subset.
Please refer to this [repo](https://github.com/harvey-fin/absence-bench) for scripts and more information.

## Citation
If you find this work useful, please cite our paper:
```Bibtex
@misc{fu2025absencebenchlanguagemodelscant,
      title={AbsenceBench: Language Models Can't Tell What's Missing}, 
      author={Harvey Yiyun Fu and Aryan Shrivastava and Jared Moore and Peter West and Chenhao Tan and Ari Holtzman},
      year={2025},
      eprint={2506.11440},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.11440}, 
}
```