anocha-mcomp commited on
Commit
db95892
·
verified ·
1 Parent(s): 9d74c02

Update model card from Confluence: Gemma-SEA-LION-v4-27B-VL

Browse files
Files changed (1) hide show
  1. gemma-sea-lion-v4-27b-vl.md +139 -0
gemma-sea-lion-v4-27b-vl.md ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Gemma-SEA-LION-v4-27B-VL
2
+
3
+ *Written by Peerat / Last updated: 2025-10-16*
4
+
5
+ **SEA-VLM** is an instruct-tuned vision-text model for the Southeast Asia (SEA) region.
6
+
7
+ Gemma-SEA-LION-v4-27B-VL has undergone post-training using instruction-image pairs datasets in Burmese, English, Indonesian, Khmer, Lao, Malay, Mandarin, Tagalog, Tamil, Thai and Vietnamese, comprising approximately 540k samples in total, to create *Gemma-SEA-LION-v4-27B-VL*.
8
+
9
+ Gemma-SEA-LION-v4-27B-VL inherits Gemma 3's:
10
+
11
+ - Large 128K context length
12
+ - Image and text understanding capabilities, including document comprehension, visual Q&A, and image-grounded reasoning
13
+ - Advanced function calling and structured outputs to allow for seamless integration into larger systems
14
+
15
+ ## Model Details
16
+
17
+ ### Model Description
18
+
19
+ We performed post-training in English and SEA languages on Gemma-SEA-LION-v4-27B-IT, a decoder model using the Gemma 3 architecture, to create Gemma-SEA-LION-v4-27B-VL.
20
+
21
+ For tokenization, the model employs the default tokenizer used in Gemma 3 27B IT.
22
+
23
+ - **Developed by:** SEACrowd and Products Pillar, AI Singapore
24
+ - **Funded by:** Singapore NRF
25
+ - **Shared by:** SEACrowd and Products Pillar, AI Singapore
26
+ - **Model type:** Decoder
27
+ - **Context length:** 128k tokens
28
+ - **Language(s) (NLP):** Burmese, English, Indonesian, Khmer, Lao, Malay, Mandarin, Tagalog, Tamil, Thai and Vietnamese
29
+ - **License:** [Gemma Terms of Use](https://ai.google.dev/gemma/terms)
30
+ - **Finetuned from model:** [Gemma-SEA-LION-v4-27B-IT](https://huggingface.co/aisingapore/Gemma-SEA-LION-v4-27B-IT)
31
+
32
+ As of 15 October 2025, Gemma-SEA-LION-v4-27B-VL excels at Southeast Asian (SEA) tasks when compared to other open models with fewer than 200 billion parameters and demonstrates performance comparable to that of larger and top closed models. For detailed rankings, please refer to the [leaderboard](https://leaderboard.sea-lion.ai/).
33
+
34
+ ## Uses
35
+
36
+ ### Out-of-Scope Use
37
+
38
+ The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes.
39
+
40
+ ## Bias, Risks, and Limitations
41
+
42
+ *The model was not tested for robustness against adversarial prompting.* It is important for users to be aware that our model exhibits certain limitations that warrant consideration. Like many LLMs, the model can hallucinate and occasionally generates irrelevant content, introducing fictional elements that are not grounded in the provided context. Users should also exercise caution in interpreting and validating the model's responses due to the potential inconsistencies.
43
+
44
+ **Limitations**
45
+
46
+ In terms of text capability, Gemma-SEA-LION-v4-27B-VL has been trained and fine-tuned exclusively on the vision-text backend. As a result, its text capabilities are expected to be comparable to those of Gemma-SEA-LION-v4-27B-IT, and may not exhibit significant improvements or differences in this area.
47
+
48
+ ## How to Get Started with the Model
49
+
50
+ Use the code below to get started with the model using the 🤗 Transformers library.
51
+
52
+ ```
53
+ from transformers import pipeline
54
+ import torch
55
+
56
+ pipe = pipeline(
57
+ "image-text-to-text",
58
+ model="aisingapore/SEA-VLM-Gemma-27B-SEALION-v1",
59
+ device="cuda",
60
+ torch_dtype=torch.bfloat16
61
+ )
62
+
63
+ messages = [
64
+ {
65
+ "role": "system",
66
+ "content": [{"type": "text", "text": "You are a helpful assistant."}]
67
+ },
68
+ {
69
+ "role": "user",
70
+ "content": [
71
+ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
72
+ {"type": "text", "text": "What animal is on the candy?"}
73
+ ]
74
+ }
75
+ ]
76
+
77
+ output = pipe(text=messages, max_new_tokens=200)
78
+ print(output[0]["generated_text"][-1]["content"])
79
+
80
+ ```
81
+ ## Training Details
82
+
83
+ ### Training Data
84
+
85
+ The dataset comprises vision-text paired in Burmese, English, Indonesian, Khmer, Lao, Malay, Mandarin, Tagalog, Tamil, Thai and Vietnamese languages, collected from a mixture of sources including web data, code, open-source datasets.
86
+
87
+ ### Training Procedure
88
+
89
+ #### Training Hyperparameters
90
+
91
+ - **Training regime:**
92
+
93
+ We perform SFT using 540k of vision-text samples written in 10 languages. Then, we perform model merging with Gemma3-27B-IT to preserve general vision-text knowledge.
94
+
95
+ ## Evaluation
96
+
97
+ ### Testing Data
98
+
99
+ To assess its cultural and visual understanding, we evaluated the model on two types of tasks using datasets focused on Southeast Asian examples:
100
+
101
+ - **Visual Question Answering (VQA):** We used Multiple Choice Question (MCQ) style VQA tasks, including MARVL, CVQA, and WorldCuisines.
102
+ - **Image Captioning:** We used the XM3600 dataset, evaluating only on examples relevant to SEA.
103
+
104
+ ### Metrics
105
+
106
+ The following metrics were used to measure performance:
107
+
108
+ - **Normalized accuracy** was the primary metric for the VQA tasks (CVQA, MARVL, and WorldCuisines).
109
+ - **RefCLIP Score** was used for the XM3600 image captioning task.
110
+
111
+ ### Results
112
+
113
+ For details on Gemma-SEA-LION-v4-27B-VL performance, please refer to ~~the SEA-HELM leaderboard,~~ [~~Leaderboard results on SEA-HELM~~](https://leaderboard.sea-lion.ai/).
114
+
115
+ ## Environmental Impact
116
+
117
+ Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
118
+
119
+ - **Hardware Type:** Nvidia H200 140GB GPUs
120
+ - **Hours used:** 13 hrs
121
+ - **Cloud Provider:** SMC H200
122
+ - **Compute Region:** Singapore
123
+ - **Carbon Emitted:** appx. 27 kg CO2 e
124
+
125
+ ## More Information
126
+
127
+ This is the repository for the commercial instruction-tuned model. The model has not been aligned for safety. Developers and users should perform their own safety fine-tuning and related security measures. In no event shall the authors be held liable for any claims, damages, or other liabilities arising from the use of the released weights and codes.
128
+
129
+ AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of the National Research Foundation or the National University of Singapore.
130
+
131
+ For more info, please contact us at [sealion@aisingapore.org](mailto:sealion@aisingapore.org), [seacrowd.research@gmail.com](mailto:seacrowd.research@gmail.com).
132
+
133
+ ## Contact
134
+
135
+ [sealion@aisingapore.org](mailto:sealion@aisingapore.org), [seacrowd.research@gmail.com](mailto:seacrowd.research@gmail.com)
136
+
137
+ ## Team
138
+
139
+ SEACrowd team members, and AI Singapore: Ahn Jeongmi, Antonyrex Sajeban, Chan Hok Teng Adwin, Cheng Zi Yi Nicholas, Choa Hsueh Mei Esther, Heng Jonathan, Huang Yuli, Hulagadri Adithya Venkatadri, Jann Railey Estrada Montalan, Kang Siow Wei Bryan, Lau Wayne, Lee Chwan Ren, Leong Wai Yi, Leong Wei Qi, Limkonchotiwat Peerat, Muhammad Ridzuan Bin Mokhtar, Nagarajan Karthik, Ng Boon Cheong Raymond, Ngee Chia Tai, Ngui Jian Gang, Nguyen Thanh Ngan, Ong Jin Jie Brandon, Ong Tat-Wee David, Ong Zhi Hao, Pereira Mark, Rengarajan Hamsawardhini, Susanto Yosephine, Sutaveephamochanon Anocha, Tan Choon Meng, Tan Chor Phin Evelyn, Tan Siao Wei Jessica, Tan Yixian, Tee Jun Yun, Teng Kok Wai Walter, Teo Eng Sipp Leslie, Tjhi William, Yeo Yeow Tong, Yong Xianbin, Zhang Zhou, Liew Rachel, Liu Bing Jie Darius