File size: 1,520 Bytes
016b5ad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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

# Jerjes/neuro-specter2-sample-data

This dataset contains anchor papers with their top-K most similar (positive) and most dissimilar (negative) papers based on SPECTER2 embeddings.

## Dataset Structure

Each row contains:
- `anchor_id`: Unique identifier for the anchor paper
- `anchor_title`: Title of the anchor paper  
- `anchor_abstract`: Abstract of the anchor paper
- `positive_pool`: List of 5 most similar papers, each as [id, title, abstract]
- `negative_pool`: List of 5 most dissimilar papers, each as [id, title, abstract]

## Dataset Statistics

- **Total anchors**: 288
- **Positives per anchor**: 5
- **Negatives per anchor**: 5
- **Embedding model**: allenai/specter2_base

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("Jerjes/neuro-specter2-sample-data")

# Access a sample
sample = dataset["train"][0]
print(f"Anchor: {sample['anchor_title']}")
print(f"Top positive: {sample['positive_pool'][0][1]}")  # title of most similar paper
print(f"Top negative: {sample['negative_pool'][0][1]}")  # title of most dissimilar paper
```

## Citation

If you use this dataset, please cite the original SPECTER2 paper:

```
@inproceedings{specter2,
    title={SPECTER2: Better Scientific Paper Representations Through Augmented Word Embeddings},
    author={Pradeep Dasigi and Kyle Lo and Iz Beltagy and Arman Cohan and Noah A. Smith and Matt Gardner},
    booktitle={Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing},
    year={2021}
}
```