File size: 4,623 Bytes
fe7aeac
 
 
 
 
 
 
c59b3c2
 
 
 
 
 
fe7aeac
 
c59b3c2
 
fe7aeac
c59b3c2
 
 
 
fe7aeac
 
 
 
 
 
 
 
299fa3c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8207041
299fa3c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
22396fe
299fa3c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
22396fe
299fa3c
22396fe
299fa3c
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
dataset_info:
  features:
  - name: title
    dtype: string
  - name: keywords
    sequence: string
  - name: messages
    list:
    - name: content
      dtype: string
    - name: role
      dtype: string
  splits:
  - name: train
    num_bytes: 1851098097.8341584
    num_examples: 145064
  - name: test
    num_bytes: 78063099.39124106
    num_examples: 6653
  download_size: 626249553
  dataset_size: 1929161197.2253995
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
---
# structured_paper_summarization

A **151 k‑example** dataset of chat‐style *prompt → structured abstract* pairs, built from ~19 000 research papers across business, management, information‑systems and social‑science domains.  Each example shows the full paper (body text) being summarised into a five‑section Emerald‑style structured abstract (Purpose, Design/methodology/approach, Findings, Practical implications, Originality/value).

---
## Why this dataset?
Large‑language models (LLMs) frequently struggle to:
1. **Condense long scientific prose** into factual, concise summaries.
2. **Follow rigid output structures** (e.g. subsection headings).

This dataset targets both challenges simultaneously, enabling fine‑tuning or instruction‑tuning of LLMs that must output *structured* scholarly abstracts.

---
## At a glance
| Split | Rows | Size (compressed) |
|-------|------|-------------------|
| train | **145 067** | 626 MB |
| test  | **6 650** |  29 MB |
| **Total** | **151 717** | ≈655 MB |

<sup>Counts taken from the Hugging Face viewer on 2025‑04‑29.</sup>

---
## Data schema
```text
{
  title:    string                    # Paper title
  keywords: list[string]             # Author‑supplied keywords (0‑23)
  messages: list[dict] length ≥ 2    # ChatML‑style conversation
}
```
### `messages` format
Each list contains alternating dictionaries with:
- `role`: either `"user"` or `"assistant"`.
- `content`: UTF‑8 text.

Typical pattern (2 items):
```jsonc
[
  {
    "role": "user",
    "content": "Summarize the following paper into structured abstract.\n\n<full paper text>"
  },
  {
    "role": "assistant",
    "content": "Purpose: …\nDesign/methodology/approach: …\nFindings: …\nPractical implications: …\nOriginality/value: …"
  }
]
```
Some papers are longer and may be truncated to ~8 k tokens.

---
## Loading the data
```python
from datasets import load_dataset

ds_train = load_dataset(
    "Neooooo/structured_paper_summarization", split="train"
)
print(ds_train[0]["messages"][1]["content"][:500])
```
The dataset is stored as Apache **Parquet** with streaming support; the example above requires ~5 s to start iterating with no local download.

---
## Suggested use‑cases
* **Instruction‑tuning** chat LLMs for long‑document summarisation.
* Research on **controlled text generation** and output formatting.
* Training **retrieval‑augmented systems** that must cite sections of the source paper.

---
## Source & construction
1. Full‑text articles were collected via institutional access to the *Emerald Insight* corpus (open‑access + subscription).
2. The canonical *structured abstract* supplied by each journal was extracted as ground truth.
3. The article’s main body was embedded into a prompt of the form shown above.
4. Data were converted to Hugging Face `datasets` ➜ auto‑parquet.

No additional manual cleaning was performed; typos and OCR artefacts may persist.

---
## Licensing & acceptable use
The article texts are **copyright their original publishers/authors** and are redistributed here *solely for non‑commercial research*.  By using this dataset you agree to:
- **Not** redistribute the raw paper texts.
- Cite the original articles in any derivative work.
- Abide by Emerald’s usage policy and your local copyright laws.

The **metadata & structured abstracts** are released under **CC BY‑NC 4.0**.  For commercial licensing, please contact the original rights‑holders.

---
## Citation
If you use this dataset, please cite:
```text
@dataset{hu_2025_structured_prompts,
  author       = {Xingyu Hu},
  title        = {structured_paper_summarization},
  year         = 2025,
  url          = {https://huggingface.co/datasets/Neooooo/structured_paper_summarization},
  note         = {Version 1.0}
}
```

---
## Contributions
Feel free to open PRs to:
- Fix metadata errors.
- Provide additional splits (validation, domain‑specific subsets).
- Add scripts for evaluation or preprocessing.

---
*Happy summarising!*