File size: 5,323 Bytes
590aade
 
a901bcb
 
 
 
 
 
 
 
 
 
 
 
590aade
a901bcb
88b58dd
a901bcb
 
33a156a
a901bcb
33a156a
ef3dd66
70ebb59
499bb9f
a901bcb
 
 
33a156a
e58c480
a901bcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
33a156a
a901bcb
 
 
ec0805e
848bd8d
 
ec0805e
a901bcb
 
ec0805e
a901bcb
b4c2b36
a901bcb
 
ec0805e
f6c9a3e
 
b4c2b36
f6c9a3e
 
a901bcb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
library_name: transformers
language:
- vi
- en
base_model:
- 5CD-AI/Vintern-1B-v2
datasets:
- vidore/colpali_train_set
- 5CD-AI/Viet-Doc-VQA
- 5CD-AI/Viet-OCR-VQA
- 5CD-AI/Viet-Doc-VQA-II
tags:
- colpali
---
<div align="center">
  <img src="colvintern.png" width="400"/>
</div>

## ColVintern-1B-v1 🇻🇳 ❄️ - Colpali version for Vietnamese.

**What's new in ColVintern-1B-v1!**
- We coded and successfully trained the **Colpali pipeline** for **Vintern-1B-v2**. The model supports RAG by extracting embedding vectors for questions and images containing related information.
- This is the first experimental version, trained on the [**Colpali dataset**](https://huggingface.co/datasets/vidore/colpali_train_set) for English and **5%** of the image-based question-answer pairs we have for Vietnamese.
- The model achieves results nearly equivalent to Colpali version 1, with strong support for Vietnamese texts and only 1 billion parameters compared to current 2B-3B Colpali models.

## Colpali Benchmarks

We tested on the [**ViDoRe benchmark**](https://huggingface.co/collections/vidore/vidore-benchmark-667173f98e70a1c0fa4db00d) from the Colpali paper. The **TabF** and **Shift** test datasets were not used because they are in French. We plan to expand to multiple languages in the near future.

|                              | ArxivQ | DocQ | InfoQ | TATQ | AI   | Energy | Gov.  | Health. | Avg.   |
|:------------------------------:|:--------:|:------:|:-------:|:------:|:------:|:--------:|:-------:|:---------:|:--------:|
| **Unstructured** Text only        |        |       |       |   |     |      |    |    |        |
| - BM25                        | -      | 34.1 | -     | 44.0 | 90.4 | 78.3   | 78.8  | 82.6    | -      |
| - BGE-M3                      | -      | 28.4 | - | 36.1 | 88.4 | 76.8 | 77.7 | 84.6 | -      |
| **Unstructured** + OCR          |        |      |       |      |      |        |       |         |     |
| - BM25                        | 31.6   | 36.8 | 62.9  | 62.7 | 92.8 | 85.9   | 83.9  | 87.2    | 68.0   |
| - BGE-M3                      | 31.4   | 25.7 | 60.1 | 50.5 | 90.2 | 83.6 | 84.9 | 91.1 |  64.7   |   
| **Unstructured** + Captioning     |        |      |       |      |      |        |       |         |  |
| - BM25                        | 40.1   | 38.4 | 70.0  | 61.5 | 88.0 | 84.7   | 82.7  | 89.2    |   69.3    |
| - BGE-M3                      | 35.7   | 32.9 | 71.9 | 43.8 | 88.8 | 83.3 | 80.4 | 91.3 |   66.0    |
| **Contrastive VLMs**                   |        |      |       |      |      |        |       |         |        |
| - Jina-CLIP                   | 25.4   | 11.9 | 35.5  | 3.3 | 15.2 | 19.7   | 21.4  | 20.8    | 19.2   |
| - Nomic-vision                | 17.1   | 10.7 | 30.1  | 2.7 | 12.9 | 10.9   | 11.4  | 15.7    | 13.9   |
| - SigLIP (Vanilla)            | 43.2   | 30.3 | 64.1  | 26.2 | 62.5 | 65.7   | 66.1  | 79.1    | 54.7   |
| **Colpali**                               |        |      |       |      |      |        |       |         |        |
| - SigLIP (Vanilla)            | 43.2   | 30.3 | 64.1  | 26.2 | 62.5 | 65.7   | 66.1  | 79.1    | 54.7   |
| - BiSigLIP (+fine-tuning)     | 58.5   | 32.9 | 70.5  | 30.5 | 74.3 | 73.7 | 74.2 | 82.3 | 62.1 |
| - BiPali (+LLM)               | 56.5   | 30.0 | 67.4  | 33.4 | 71.2 | 61.9 | 73.8 | 73.6 | 58.5 |
| - ColPali (+Late Inter.)      | **79.1**   | **54.4** | 81.8  | **65.8** | **96.2** | **91.0** | **92.7** | 94.4 | **81.3** |
| **Ours**                               |        |      |       |      |      |        |       |         |        |
| - ColVintern-1B (+Late Inter.)  | 71.6   | 48.3 | **84.6**  | 59.6 | 92.9 | 88.7 | 89.4 | **95.2** | 78.8 |

We are expanding the training dataset for upcoming versions, including adding hard negative mining techniques, increasing GPU VRAM, etc., to achieve better results.

## Examples

Input Images:
<div style="display: flex; gap: 20px;">
  <img src="ex1.jpg" width="300"/>
  <img src="ex2.jpg" width="300"/>
</div>

Input Queries:
```
queries = ["Cảng Hải Phòng thông báo gì ?","Phí giao hàng bao nhiêu ?"]
```

Output Scores:
| Query                              | Image 1 Score | Image 2 Score |
|------------------------------------|---------------|---------------|
| Cảng Hải Phòng thông báo gì ? | 62.50   | 60.00         |
| Phí giao hàng bao nhiêu ?          | 60.75         | 62.75         |


## Quickstart


```python
import torch
from PIL import Image
from transformers import AutoModel, AutoTokenizer, AutoProcessor

model_name = "5CD-AI/ColVintern-1B-v1"

processor =  AutoProcessor.from_pretrained(
    model_name,
    trust_remote_code=True
)
model = AutoModel.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True,
).eval().cuda()

```

## Citation 

<!-- ```
@misc{doan2024vintern1befficientmultimodallarge,
      title={Vintern-1B: An Efficient Multimodal Large Language Model for Vietnamese}, 
      author={Khang T. Doan and Bao G. Huynh and Dung T. Hoang and Thuc D. Pham and Nhat H. Pham and Quan T. M. Nguyen and Bang Q. Vo and Suong N. Hoang},
      year={2024},
      eprint={2408.12480},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2408.12480}, 
}
``` -->