Add dataset card
Browse files
README.md
ADDED
@@ -0,0 +1,72 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
language:
|
3 |
+
- en
|
4 |
+
license: mit
|
5 |
+
tags:
|
6 |
+
- two-tower
|
7 |
+
- semantic-search
|
8 |
+
- document-retrieval
|
9 |
+
- information-retrieval
|
10 |
+
- dual-encoder
|
11 |
+
---
|
12 |
+
|
13 |
+
# mlx7-two-tower-data
|
14 |
+
|
15 |
+
This repository contains datasets used for training Two-Tower (Dual Encoder) models for document retrieval.
|
16 |
+
|
17 |
+
## Dataset Description
|
18 |
+
|
19 |
+
The datasets provided here are structured for training dual encoder models with various sampling strategies:
|
20 |
+
|
21 |
+
- **classic_triplets**: 48.2 MB
|
22 |
+
- **intra_query_neg**: 47.6 MB
|
23 |
+
- **multi_pos_multi_neg**: 126.5 MB
|
24 |
+
|
25 |
+
### Dataset Details
|
26 |
+
|
27 |
+
- **classic_triplets.parquet**: Standard triplet format with (query, positive_document, negative_document)
|
28 |
+
- **intra_query_neg.parquet**: Negative examples selected from within the same query batch
|
29 |
+
- **multi_pos_multi_neg.parquet**: Multiple positive and negative examples per query
|
30 |
+
|
31 |
+
## Usage
|
32 |
+
|
33 |
+
```python
|
34 |
+
import pandas as pd
|
35 |
+
|
36 |
+
# Load a dataset
|
37 |
+
df = pd.read_parquet("classic_triplets.parquet")
|
38 |
+
|
39 |
+
# View the schema
|
40 |
+
print(df.columns)
|
41 |
+
|
42 |
+
# Example of working with the data
|
43 |
+
queries = df["q_text"].tolist()
|
44 |
+
positive_docs = df["d_pos_text"].tolist()
|
45 |
+
negative_docs = df["d_neg_text"].tolist()
|
46 |
+
```
|
47 |
+
|
48 |
+
## Data Source and Preparation
|
49 |
+
|
50 |
+
These datasets are derived from the MS MARCO passage retrieval dataset, processed to create effective training examples for two-tower models.
|
51 |
+
|
52 |
+
## Dataset Structure
|
53 |
+
|
54 |
+
The datasets follow a common schema with the following fields:
|
55 |
+
- `q_text`: Query text
|
56 |
+
- `d_pos_text`: Positive (relevant) document text
|
57 |
+
- `d_neg_text`: Negative (non-relevant) document text
|
58 |
+
|
59 |
+
Additional fields may be present in specific datasets.
|
60 |
+
|
61 |
+
## Citation
|
62 |
+
|
63 |
+
If you use this dataset in your research, please cite the original MS MARCO dataset:
|
64 |
+
|
65 |
+
```
|
66 |
+
@article{msmarco,
|
67 |
+
title={MS MARCO: A Human Generated MAchine Reading COmprehension Dataset},
|
68 |
+
author={Nguyen, Tri and Rosenberg, Matthew and Song, Xia and Gao, Jianfeng and Tiwary, Saurabh and Majumder, Rangan and Deng, Li},
|
69 |
+
journal={arXiv preprint arXiv:1611.09268},
|
70 |
+
year={2016}
|
71 |
+
}
|
72 |
+
```
|