File size: 4,745 Bytes
0e23c8c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
bb475dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0e23c8c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
170
171
172
173
174
175
176
177
---
license: cc-by-4.0
task_categories:
- text-classification
- sentence-similarity
- feature-extraction
language:
- en
tags:
- biomedical
- scientific-literature
- pubmed
- pmc
- embeddings
- soda-vec
size_categories:
- 10M<n<100M
dataset_info:
  features:
  - name: pmid
    dtype: string
  - name: title
    dtype: string
  - name: abstract
    dtype: string
  - name: doi
    dtype: string
  splits:
  - name: train
    num_bytes: 39688553174
    num_examples: 26473900
  - name: validation
    num_bytes: 74918135
    num_examples: 50000
  - name: test
    num_bytes: 74931494
    num_examples: 50000
  download_size: 23525241784
  dataset_size: 39838402803
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---

# SODA-VEC Clean Dataset

This is a **cleaned and filtered version** of the SODA-VEC dataset, containing high-quality biomedical title-abstract pairs from PubMed Central (PMC) articles.

## Dataset Overview

- **Total examples**: 26,573,900
- **Training set**: 26,473,900 examples (99.6%)
- **Validation set**: 50,000 examples (0.2%)
- **Test set**: 50,000 examples (0.2%)

## Quality Filtering Applied

This dataset has been processed with the following quality filters:

### Abstract Length Filtering
- **Minimum length**: 128 characters
- **Maximum length**: 6,000 characters
- **Rationale**: Removes fragments and overly long texts while preserving scientific abstracts

### Retention Statistics
- **Original dataset**: ~26.6M examples
- **After filtering**: 26,573,900 examples
- **Retention rate**: ~99.7%

### Content Statistics (sample):
- **Title length**: ~100 ± 50 chars
- **Abstract length**: ~1,300 ± 600 chars
- **Title range**: 10-500 chars
- **Abstract range**: 128-6,000 chars

## Length Distributions

![Length Distributions](length_distributions.png)

The dataset shows well-balanced length distributions:
- **Title lengths**: Centered around 100 characters with good variance
- **Abstract lengths**: Normally distributed around 1,300 characters
- **Quality filtering**: Clearly removes outliers while preserving natural variation

![Length Statistics](length_statistics.png)

## Data Fields

Each example contains:

- **`title`** (string): The title of the scientific article
- **`abstract`** (string): The abstract of the scientific article
- **`pmcid`** (string): PubMed Central ID for the article

## Intended Use

This dataset is designed for:

### Primary Applications
- **Scientific text embeddings**: Training domain-specific embedding models
- **Biomedical NLP**: Fine-tuning language models on scientific literature
- **Semantic similarity**: Learning representations for scientific text matching
- **Information retrieval**: Building search systems for biomedical literature

### Research Applications
- Representation learning for scientific texts
- Cross-modal learning (title-abstract relationships)
- Domain adaptation for biomedical language models
- Evaluation of scientific text understanding systems

## Data Source & Methodology

### Original Dataset
Based on the SODA-VEC dataset from PubMed Central articles.

### Processing Pipeline
1. **Data loading**: Combined train/validation/test splits from original dataset
2. **Quality filtering**: Applied length-based filters to ensure high-quality pairs
3. **Split creation**: Created new balanced train/validation/test splits
4. **Validation**: Verified data integrity and distribution balance

### Quality Assurance
- Length distribution analysis
- Duplicate detection and removal
- Content quality validation
- Statistical validation of splits

## Usage Example

```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("EMBO/soda-vec-data-full_pmc_title_abstract")

# Access different splits
train_data = dataset["train"]
val_data = dataset["validation"] 
test_data = dataset["test"]

# Example usage
for example in train_data.take(1):
    print(f"Title: {example['title']}")
    print(f"Abstract: {example['abstract']}")
    print(f"PMC ID: {example['pmcid']}")
```

## Citation

If you use this dataset in your research, please cite the original SODA-VEC paper:

```bibtex
@article{soda-vec-2024,
  title={SODA-VEC: Training Vector Representations of Scientific Literature},
  author={...},
  journal={...},
  year={2024}
}
```

## License

This dataset is released under the **CC-BY-4.0** license, consistent with PubMed Central's open access requirements.

## Contact

For questions about this dataset, please contact the EMBO team or open an issue in the dataset repository.

---

*Dataset processed with quality filters and balanced splits for optimal training performance.*