Hugging Face's logo

FreedomIntelligence
/
ALLaVA-Phi2-2_7B

Text Generation
Transformers
Safetensors
English
llava_phi
custom_code
Model card Files Community
g-h-chen commited on
Commit
fd86fe6
β€’
1 Parent(s): c060d67

upload README.md

Browse files
Files changed (1) hide show
  1. README.md +152 -3
README.md CHANGED
@@ -1,3 +1,152 @@
1
- ---
2
- license: cc-by-4.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ datasets:
4
+ - FreedomIntelligence/ALLaVA-4V
5
+ language:
6
+ - en
7
+ pipeline_tag: text-generation
8
+ ---
9
+
10
+
11
+ # ALLaVA: Harnessing GPT4V-synthesized Data for A Lite Vision-Language Model
12
+
13
+
14
+
15
+ <p align="center">
16
+ ⚑ALLaVA is a project that provides a large-scale GPT4V-synthesized dataset for training LVLMs.⚑
17
+ </p>
18
+
19
+ <!-- <p align="center">
20
+
21
+ ![Python 3.10](https://img.shields.io/badge/Python-3.10-lightblue) ![Pytorch 1.13.0](https://img.shields.io/badge/PyTorch-2.1.1-lightblue) ![transformers](https://img.shields.io/badge/transformers-4.37.0-lightblue)
22
+ </p> -->
23
+
24
+
25
+
26
+ <p align="center">
27
+ πŸ“ƒ <a href="https://arxiv.org/abs/2402.11684" target="_blank">Paper</a> β€’ 🌐 <a href="https://allava.freedomai.cn/#/" target="_blank">Demo</a> β€’ πŸ‘¨πŸ»β€πŸ’» <a href="https://github.com/FreedomIntelligence/ALLaVA" target="_blank">Github</a>
28
+ </p>
29
+
30
+ <p align="center">
31
+ πŸ€— <a href="https://huggingface.co/datasets/FreedomIntelligence/ALLaVA-4V" target="_blank">ALLaVA-4V Dataset</a>
32
+ </p>
33
+
34
+ <p align="center">
35
+ πŸ€— <a href="https://huggingface.co/FreedomIntelligence/ALLaVA-Phi3-mini-128k" target="_blank">ALLaVA-Phi3-mini-128k</a>
36
+ β€’ πŸ€— <a href="https://huggingface.co/FreedomIntelligence/ALLaVA-StableLM2-1_6B" target="_blank">ALLaVA-StableLM2-1_6B</a>
37
+ β€’ πŸ€— <a href="https://huggingface.co/FreedomIntelligence/ALLaVA-Phi2-2_7B" target="_blank">ALLaVA-Phi2-2_7B</a>
38
+ </p>
39
+
40
+ <!-- <p align="center">
41
+ πŸ“ƒ <a href="https://arxiv.org/abs/2402.11684" target="_blank">Paper</a> β€’ 🌐 <a href="https://allava.freedomai.cn/#/" target="_blank">Demo</a> β€’ πŸ€— <a href="https://huggingface.co/datasets/FreedomIntelligence/ALLaVA-4V" target="_blank">ALLaVA-4V Dataset</a> β€’ πŸ€— <a href="https://huggingface.co/FreedomIntelligence/ALLaVA-3B-Longer" target="_blank">ALLaVA-3B-Longer</a> β€’ πŸ€— <a href="https://huggingface.co/FreedomIntelligence/ALLaVA-3B" target="_blank">ALLaVA-3B</a>
42
+ <br> <a href="https://github.com/FreedomIntelligence/CMB/blob/main/README_zh.md"> δΈ­ζ–‡</a> | <a href="https://github.com/FreedomIntelligence/CMB/blob/main/README.md"> English
43
+ </p> -->
44
+
45
+ ## Benchmark Result
46
+
47
+ Our models [**ALLaVA-Phi3-mini-128k**](https://huggingface.co/FreedomIntelligence/ALLaVA-Phi3-mini-128k),
48
+ [**ALLaVA-StableLM2-1_6B**](https://huggingface.co/FreedomIntelligence/ALLaVA-StableLM2-1_6B)
49
+ and [**ALLaVA-Phi2-2_7B**](https://huggingface.co/FreedomIntelligence/ALLaVA-Phi2-2_7B)
50
+ achieve competitive results on 17 benchmarks.
51
+
52
+
53
+ | Models | Vicuna-80 | GQA | HallusionBench | MME-P | MMVP | TouchStone | TextVQA | MME-C | MathVista | MM-Vet | MMMU-val | SQA (img) | LLaVA (In-the-Wild) | MLLM-Bench | MMB-en | MMB-cn | SEEDBench (img, v1) |
54
+ |---------------------------|-----------|-----|-------|-------|------|----|---------|-------|----|--------|-----------------|---------|---------------|----|--------|--------|--------------------|
55
+ | **Large VLMs** | | | | | | | | | | | | | | | | | |
56
+ | BLIP-2 | - | - | - | - | - | - | - | - | - | 22.4 | 34.4 | - | - | 3.0*| - | - | 49.7 |
57
+ | InstructBLIP | - | 49.5| - | - | - | - | - | - | - | 25.6 | - | - | 58.2 | - | 44.0 | - | - |
58
+ | Qwen-VL-Chat | - | 57.5| - | 1487.6| - | - | 61.5 | 360.7 | - | 31.1 | - | 68.2 | - | - | 60.6 | 56.7 | 65.4 |
59
+ | LLaVA-1.5-7B | 13.8* | 62.0| 36.6* | 1504.4*| 24.7*| 594.9*| 58.2| 324.6*| 25.0*| 31.1| 35.1*| 66.8| 65.4| 23.0*| 64.3| 58.3| 66.1|
60
+ | LLaVA-1.5-13B | 22.5 | 63.3| 36.5* | 1531.3 | 38.0*| 617.7*| 61.3| 295.4| 28.3*| 35.4| 34.4*| 71.6| 72.5| -| 67.7| 63.6| 68.2|
61
+ | LVIS-7B | - | 62.6| - | - | - | - | 58.7 | - | - | 31.5 | - | - | 67.0 | 29.0*| 66.2 | - | - |
62
+ | LVIS-13B | - | 63.6*| - | - | - | - | 62.5* | - | - | 37.4* | - | - | 71.3* | - | 68.0* | - | - |
63
+ | ShareGPT4V-7B | 13.8* | 63.3| 36.0* | 1540.1*| 34.0*| 637.2*| 60.4| 346.1*| 24.7*| 37.6| 35.4*| 68.4*| 72.6| 30.2*| 68.8| 61.0*| 69.7|
64
+ | ShareGPT4V-13B | 17.5* | 64.8| 39.0* | 1576.1*| 35.3*| 648.7*| 62.2| 309.3*| 28.8*| 43.1| 35.6*| 70.0*| 79.9| 35.5*| 71.2| 61.7*| 70.8|
65
+ | **4B-scale Lite VLMs** | | | | | | | | | | | | | | | | | |
66
+ | MobileVLM-v2 | 5.0* | 61.1| 30.8* | 1440.5 | 18.7*| 541.0*| 57.5| 261.8*| 28.3*| 26.1*| 30.8*| 70.0| 53.2*| 15.7*| 63.2| 43.2*| 64.5*|
67
+ | Mipha-3B | 16.2* | **63.9**| 34.3*| **1488.9**| 32.0*| 619.0*| 56.6| 285.0*| 27.8*| 33.5*| 35.8*| 70.9| 64.7*| 23.1*| **69.7**| 42.9*| **71.2***|
68
+ | TinyLLaVA | 15.6* | 62.1| 37.2* | 1465.5*| 33.3*| 663.5*| **60.3**| 281.1*| 30.3*| 37.5| 38.4| **73.0**| 70.8*| 29.8*| **69.7***| 42.8*| 70.4*|
69
+ | **Ours** | | | | | | | | | | | | | | | | | |
70
+ | **ALLaVA-Phi2** | 49.4 | 48.8| 24.8 | 1316.2| **36.0**| 632.0| 49.5| 301.8| 27.4| 32.2| 35.3| 67.6| 69.4| 43.6| 64.0| 40.8| 65.2|
71
+ | **ALLaVA-StableLM2** | 38.8 | 49.8| 25.3 | 1311.7| 34.0 | 655.2| 51.7| 257.9| 27.7| 31.7| 33.3| 64.7| **72.0**| 39.3| 64.6| 49.8| 65.7|
72
+ | **ALLaVA-Phi3** | **56.9**| 52.2| **48.1**| 1382.3| 32.7| **667.8**| 53.0| **347.1**| **32.9**| **37.8**| **41.1**| 64.0| 68.5| **54.8**| 68.1| **55.3**| 69.0|
73
+
74
+
75
+ > \* denotes the results of our evaluation. **Bold numbers** are the best results among all 4B-scale LVLMs.The detailed information of each benchmark is shown in Table 4 of our [technical report](https://arxiv.org/pdf/2402.11684.pdf).
76
+
77
+
78
+
79
+ ## 🏭 Inference
80
+
81
+ ### Load from πŸ€— (Recommended)
82
+ See the [example script](https://github.com/FreedomIntelligence/ALLaVA/blob/main/allava/serve/huggingface_inference.py).
83
+
84
+ ### CLI
85
+ See [here](https://github.com/FreedomIntelligence/ALLaVA/tree/main?tab=readme-ov-file#cli) for CLI code snippet.
86
+
87
+
88
+
89
+ ## πŸ‹οΈβ€β™‚οΈ Training
90
+
91
+ ### Data
92
+ <!-- <div align=center>
93
+ <img src="training_datasets_by_stage.jpg" width = "640" alt="training_datasets" align=center />
94
+ </div> -->
95
+
96
+ ALLaVA uses 795K and 1.4M data for PT. and FT., respectively.
97
+
98
+
99
+ ### Code
100
+ The training code is largely based on [LLaVA-v1.5](https://github.com/haotian-liu/LLaVA).
101
+ We wholeheartedly express our gratitude for their invaluable contributions to open-sourcing LVLMs.
102
+
103
+ <!-- ### Cost
104
+ We train our models on 8*A800 GPUs.
105
+ [ALLaVA-3B-Longer](https://huggingface.co/FreedomIntelligence/ALLaVA-3B-Longer) takes 8.3h for PT and 21.3h for FT.
106
+ [ALLaVA-3B](https://huggingface.co/FreedomIntelligence/ALLaVA-3B) takes 8.3h for PT and 10.6h for FT.
107
+ These two models share the same PT procedure. -->
108
+
109
+
110
+ ### Hyperparameters
111
+
112
+ | Global Batch Size| ZeRO Stage| Optimizer | Max LR| Min LR | Scheduler | Weight decay |
113
+ | ---: | ---: |--:| ---: | ---: | ---: | ---: | ---: |
114
+ | 256 (PT) / 128 (FT) | 1| AdamW | 2e-5 | 2e-6 | CosineAnnealingWarmRestarts | 0 |
115
+
116
+ The LM backbone, projector are trainable, while the vision encoder is kept frozen.
117
+ **The trainabilities of each module are the same for both stages.**
118
+
119
+
120
+ ## πŸ“š ALLaVA-4V Data
121
+
122
+ The majority part of training data is [ALLaVA-4V](https://huggingface.co/datasets/FreedomIntelligence/ALLaVA-4V). See [here](https://github.com/FreedomIntelligence/ALLaVA/tree/main?tab=readme-ov-file#data-preparation) to prepare it for training.
123
+
124
+
125
+ ## πŸ™Œ Contributors
126
+
127
+ - Project Leader: [Guiming Hardy Chen](https://g-h-chen.github.io/)
128
+
129
+ - Data: Shunian Chen, [Junying Chen](https://jymchen.github.io/), Xiangbo Wu
130
+
131
+ - Evaluation: [Ruifei Zhang](https://scholar.google.com/citations?user=W4zOhmEAAAAJ&hl=zh-CN)
132
+
133
+ - Deployment: Xiangbo Wu, Zhiyi Zhang
134
+
135
+ - Advising: [Zhihong Chen](https://zhjohnchan.github.io/), [Benyou Wang](https://wabyking.github.io/old.html)
136
+
137
+ - Others: Jianquan Li, [Xiang Wan](https://scholar.google.com/citations?user=e3_kWigAAAAJ&hl=zh-CN)
138
+
139
+
140
+
141
+
142
+
143
+ ## πŸ“ Citation
144
+ If you find our data useful, please consider citing our work! We are FreedomIntelligence from [Shenzhen Research Institute of Big Data](http://sribd.cn/en) and [The Chinese University of Hong Kong, Shenzhen](https://sds.cuhk.edu.cn/en)
145
+ ```
146
+ @article{chen2024allava,
147
+ title={ALLaVA: Harnessing GPT4V-synthesized Data for A Lite Vision-Language Model},
148
+ author={Chen, Guiming Hardy and Chen, Shunian and Zhang, Ruifei and Chen, Junying and Wu, Xiangbo and Zhang, Zhiyi and Chen, Zhihong and Li, Jianquan and Wan, Xiang and Wang, Benyou},
149
+ journal={arXiv preprint arXiv:2402.11684},
150
+ year={2024}
151
+ }
152
+ ```