Hugging Face's logo

Datasets:
Ahren09
/
SARA-QASPER

Tasks:
Question Answering
Modalities:
Text
Formats:
json
Languages:
English
Size:
10K - 100K
ArXiv:
Tags:
retrieval-augmented-generation
context-compression
scientific-papers
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
SARA-QASPER
File size: 4,364 Bytes
732f806
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
f1a2b45
732f806
 
 
 
 
 
 
 
f1a2b45
732f806
 
 
f1a2b45
 
 
 
732f806
 
 
 
 
 
 
 
 
 
 
 
 
f1a2b45
 
 
732f806
 
 
 
f1a2b45
 
732f806
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
license: cc-by-4.0
language:
- en
task_categories:
- question-answering
pretty_name: SARA QASPER (reformatted)
tags:
- retrieval-augmented-generation
- context-compression
- scientific-papers
configs:
- config_name: qa
  default: true
  data_files:
  - split: train
    path: QASPER_train.jsonl
  - split: test
    path: QASPER_test.jsonl
- config_name: compression_alignment
  data_files:
  - split: train
    path: QASPER_compression_alignment_train.jsonl
  - split: validation
    path: QASPER_compression_alignment_dev.jsonl
---

# SARA QASPER (reformatted)

Reformatted [QASPER](https://huggingface.co/datasets/allenai/qasper) data used by
**SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression** (ACL 2026,
[arXiv:2507.05633](https://arxiv.org/abs/2507.05633)). Code: [Ahren09/SARA](https://github.com/Ahren09/SARA).

The SARA Quick Start (`python -m src.data.make_qasper_splits`) downloads this dataset automatically;
you can also load it directly:

```python
from datasets import load_dataset

qa = load_dataset("Ahren09/SARA-QASPER", "qa")                          # train / test
align = load_dataset("Ahren09/SARA-QASPER", "compression_alignment")    # train / validation
```

## Configs

### `qa` (default)

One record per answerable QASPER question. `train` (2,321 rows over the official
train papers) and `test` (1,312 rows, official test papers).

| Field | Type | Description |
| :--- | :--- | :--- |
| `id` | str | Running row index. |
| `example_id` | str | Paper index — used for leakage-safe document-level train/dev splitting. |
| `question` | str | The QASPER question. |
| `context` | list[str] | BM25-ranked paper contexts, each formatted `"Section name\t<text>"`. Retrieval in SARA is built in-memory from this field; no separate index is needed. |
| `answer` | str | Short gold answer (a span, value, phrase, or yes/no) derived from the official QASPER annotations. Used as both the training target and the evaluation reference. |
| `choices` | null | Unused for QASPER (kept for schema compatibility with multiple-choice datasets). |
| `question_type` | str | One of `extractive`, `free_form`, `yes_no`, `unanswerable`. |

Models trained on this data answer with chain-of-thought followed by the final short answer wrapped
in tags: `... reasoning ... <answer>short answer</answer>`; evaluation extracts the tagged span and
scores it against `answer`.

### `compression_alignment`

Text snippets from QASPER paper bodies used for the SARA projector-alignment warm-up (the projector
learns to reconstruct a document from its semantic compression vector before QA fine-tuning).
`train` (22,111 rows) and `validation` (200 rows); each record is `{"text": "<document text snippet>"}`.

## Provenance

- Derived from [`allenai/qasper`](https://huggingface.co/datasets/allenai/qasper)
  (Dasigi et al., NAACL 2021), released under **CC BY 4.0**. This derivative is released under the
  same license with attribution.
- `context` was built from each paper's title/abstract/sections and BM25-ranked per question;
  `BIBREF` citation markers were stripped.
- Earlier revisions of this dataset carried an additional LLM-rewritten `answer_reformatted` field;
  it has been removed — the short gold `answer` is the single reference. The old revision remains
  available in this repo's git history.

## Checksums (sha256)

```
0ec3d1bbab2f85341a9432b263ca90669f5cfd45bb2163170cedddf8ff60742c  QASPER_train.jsonl
a26afe8a11e350e5a860c83b75ee4938b2b89f19dc4883cea063dd90642bc1e1  QASPER_test.jsonl
1d6386a84408127a20f01db92f8b9a73ec9499d884d0cab3c2c9f07c4494aa12  QASPER_compression_alignment_train.jsonl
9de81c3cbf42aa4b4001a57762c37f981eb2ae0c5b468c3bfebffe65438f8a0f  QASPER_compression_alignment_dev.jsonl
```

## Citation

```bibtex
@inproceedings{jin2025sara,
  title={SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression},
  author={Jin, Yiqiao and Sharma, Kartik and Rakesh, Vineeth and Dou, Yingtong and Pan, Menghai and Das, Mahashweta and Kumar, Srijan},
  booktitle={ACL},
  year={2026}
}

@inproceedings{dasigi2021dataset,
  title={A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers},
  author={Dasigi, Pradeep and Lo, Kyle and Beltagy, Iz and Cohan, Arman and Smith, Noah A. and Gardner, Matt},
  booktitle={NAACL},
  year={2021}
}
```