anocha-mcomp commited on
Commit
902f2c8
·
verified ·
1 Parent(s): bb21f0d

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +0 -247
README.md CHANGED
@@ -12,250 +12,3 @@ language:
12
  - vi
13
  ---
14
 
15
- # Uploaded model
16
-
17
- - **Developed by:** anocha-mcomp
18
- - **License:** apache-2.0
19
- - **Finetuned from model :** unsloth/llama-3.2-1b-instruct-bnb-4bit
20
-
21
- This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.
22
-
23
- Last update 17 Sep 2025
24
-
25
- [<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)
26
-
27
- ---
28
- # For reference on model card metadata, see the spec: https://github.com/huggingface/hub-docs/blob/main/modelcard.md?plain=1
29
- # Doc / guide: https://huggingface.co/docs/hub/model-cards
30
- Gemma-SEA-LION-v4-27B-IT (IT Model) Last updated: 2025-08-19
31
- ---
32
-
33
- # This is a test model card copied from aisingapore/Gemma-SEA-LION-v4-27B-IT
34
-
35
- Last updated: 2025-08-20
36
-
37
- **SEA-LION** is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned
38
- for the Southeast Asia (SEA) region.
39
-
40
-
41
-
42
- <!-- Introduction -->
43
- ## Introduction
44
-
45
-
46
-
47
- SEA-LION stands for *Southeast Asian Languages In One Network*.
48
-
49
- Gemma-SEA-LION-v4-27B is based on Gemma 3 (which supports over 100 languages) and is a multilingual model
50
- which has undergone continued pre-training on approximately **500B** tokens across 11 SEA languages:
51
- Bahasa Indonesia, Burmese, Chinese, English, Khmer, Lao, Malay, Tagalog, Tamil, Thai and Vietnamese.
52
-
53
- We continued post-training on Gemma-SEA-LION-v4-27B using a QA pairs dataset in
54
- Bahasa Indonesia, Burmese, Chinese, English, Khmer, Lao, Malay, Tagalog, Tamil, Thai, and Vietnamese
55
- to create *Gemma-SEA-LION-v4-27B-IT*.
56
-
57
- For tokenization, the model employs the default tokenizer used in Gemma 3 27B IT.
58
-
59
-
60
- At a glance:
61
-
62
- - **Developed by:** Products Pillar, AI Singapore
63
- - **Funded by:** Singapore NRF
64
- - **Shared by:** Products Pillar, AI Singapore
65
- - **Model type:** Decoder
66
- - **Context length:** 128k tokens
67
- - **Language(s):** Bahasa Indonesia, Burmese, Chinese, English, Khmer, Lao, Malay, Tagalog, Tamil, Thai and Vietnamese
68
- - **License:** [Gemma Terms of Use](https://ai.google.dev/gemma/terms)
69
- - **Finetuned from model:** Gemma-SEA-LION-v4-27B
70
-
71
- As of 25 Aug 2025, Gemma-SEA-LION-v4-27B-IT 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/).
72
-
73
-
74
- ## Training Details
75
-
76
- ### Training Data
77
-
78
- <!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
79
-
80
- The dataset comprises Bahasa Indonesia, Burmese, Chinese, English, Khmer, Lao, Malay, Tagalog, Tamil,
81
- Thai and Vietnamese languages, collected from a mixture of sources including web data, code, open-source datasets,
82
- and synthetically generated datasets, amounting to a total of 500 billion tokens.
83
-
84
-
85
- ### Training Procedure
86
-
87
- - **Training regime:** We perform post-training using a variety of Reinforcement Learning (RL) methods.
88
- The instruction fine-tuning dataset combines our SEA-Instruct, Infinity-Instruct,
89
- and OpenMath-Instruct 2 with open-source datasets such as
90
- nvidia/Llama-Nemotron-Post-Training-Dataset (RL set) and zwhe99/DeepMath-103K.
91
- Prompt sampling is guided by a gradient-based analysis process.
92
-
93
- Our post-training workflow consists of multiple stages: instruction fine-tuning,
94
- model merging, online RL for both instruction following and math using DRGPPO,
95
- and on-policy alignment via APO. For alignment, rejected-chosen pairs are generated
96
- from the target model, with the *chosen* responses obtained by rewriting and improving upon
97
- the *rejected* outputs.
98
-
99
- For the safety alignment process, we employed KTO and focused on a selected set of Southeast Asian languages.
100
- Safety alignment was conducted in several of these languages to ensure general safety across contexts.
101
- <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
102
-
103
- ## Evaluation
104
-
105
- <!-- This section describes the evaluation protocols and provides the results. -->
106
- We evaluated Gemma-SEA-LION-v4-27B-IT on both general language capabilities and instruction-following capabilites.
107
- ### Testing Data, Factors & Metrics
108
-
109
- #### Testing Data
110
-
111
- <!-- This should link to a Dataset Card if possible. -->
112
-
113
- We evaluated Gemma-SEA-LION-v4-27B-IT on general language, multi-turn chat and instruction-following capabilities.
114
-
115
- **Testing Data**
116
-
117
- General language capabilities
118
-
119
- For the evaluation of general language capabilities, we employed the [SEA-HELM evaluation benchmark](https://arxiv.org/abs/2502.14301) across a variety of tasks.
120
- These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng),
121
- Abstractive Summarisation (Abssum), Causal Reasoning (Causal), Natural Language Inference (NLI), Linguistic Diagnostics (LINDSEA), Cultural Knowledge (Kalahi)
122
- and Global MMLU Lite.
123
-
124
- Instruction-following and Multi-turn Chat
125
-
126
- We evaluated the models on instruction-following and multi-turn chat capabilities with SEA-IFEval (based on [IFEval](https://arxiv.org/abs/2311.07911)) and SEA-MTBench (based on [MT-Bench](https://arxiv.org/abs/2306.05685)) respectively.
127
- The two datasets were originally in English, the linguists and native speakers in the team worked together to filter, localise and translate the datasets into the respective target languages to ensure that the examples remained reasonable, meaningful and natural.
128
-
129
-
130
- #### Factors
131
-
132
- <!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
133
-
134
- All evaluations were run with the model specific generation parameters defined in the model config. Each evaluation comprised of 8 runs with different seeds and the final results were averaged across these runs.
135
-
136
- For all tasks, the model was expected to provide an answer tag from which the answer was automatically extracted. For tasks where options were provided, the answer should comprise one of the pre-defined options.
137
-
138
- The evaluation was done zero-shot with native prompts on a sample of 100-1000 instances for each dataset.
139
-
140
- SEA-IFEval
141
-
142
- SEA-IFEval evaluates a model's ability to adhere to constraints provided in the prompt,
143
- for example beginning a response with a specific word/phrase or answering with a certain number of sections.
144
- Additionally, accuracy is normalised by the proportion of responses in the correct language
145
- (if the model performs the task correctly but responds in the wrong language, it is judged to have failed the task).
146
-
147
- SEA-MTBench
148
-
149
- SEA-MTBench evaluates a model's ability to engage in multi-turn (2 turns) conversations and respond in ways that align with human needs.
150
- We use `gpt-4.1-2025-04-14` as the judge model and compare against `gpt-4.1-2025-04-14` as the baseline model.
151
- The metric used is the weighted win rate against the baseline model (i.e. average win rate across each category:
152
- Math, Reasoning, STEM, Humanities, Roleplay, Writing, Extraction).
153
-
154
-
155
- #### Metrics
156
-
157
- <!-- These are the evaluation metrics being used, ideally with a description of why. -->
158
-
159
- The following metrics were used:
160
- | Task | Metric |
161
- |--------------------------------------|----------------------------------------|
162
- | Sentiment Analysis | Accuracy |
163
- | Extractive QA (ID, VI, TH, TA) | ChrF++ |
164
- | MCQ-QA (TL, MY, MS) | Accuracy |
165
- | Metaphor | Accuracy |
166
- | Abstractive Summarisation | Rouge-L |
167
- | Translations | MetricX-24 score (with reference) |
168
- | Causal Reasoning | Accuracy |
169
- | Natural Language Inference | Accuracy |
170
- | LINDSEA | Accuracy |
171
- | Global MMLU Lite | Accuracy |
172
- | Kalahi | Accuracy |
173
- | SEA-IFEval | Accuracy |
174
- | SEA-MTBench | Win rate against a reference |
175
- | Toxicity Detection | Accuracy |
176
-
177
-
178
- ### Results
179
-
180
- Coming soon!
181
-
182
- For details on Gemma-SEA-LION-v4-27B-IT performance, please refer to the SEA-HELM leaderboard, [Leaderboard results on SEA-HELM](https://leaderboard.sea-lion.ai/).
183
-
184
-
185
- ## How to Get Started with the Model
186
-
187
- ## Download the Models(s)
188
-
189
- Gemma-SEA-LION-v4-27B-IT models are available for download via the following channels:
190
-
191
- 🤗[Gemma-SEA-LION-v4](https://huggingface.co/aisingapore/Gemma-SEA-LION-v4)
192
-
193
- ## Usage
194
- Use the code below to get started with the model.
195
-
196
- Use the code below to get started with the model using the 🤗 Transformers library.
197
- ```python
198
- from transformers import pipeline
199
- import torch
200
-
201
- pipe = pipeline(
202
- "text-generation",
203
- model="aisingapore/Gemma-SEA-LION-v4-27B-IT",
204
- device="cuda",
205
- torch_dtype=torch.bfloat16
206
- )
207
-
208
- messages = [
209
- {
210
- "role": "system",
211
- "content": [{"type": "text", "text": "You are a helpful assistant."}]
212
- },
213
- {
214
- "role": "user",
215
- "content": [
216
- {"type": "text", "text": "Write a poem on southeast asian countries in Indonesian."}
217
- ]
218
- }
219
- ]
220
-
221
- output = pipe(text=messages, max_new_tokens=200)
222
- print(output[0]["generated_text"][-1]["content"])
223
- ```
224
-
225
-
226
- ## Disclaimer
227
-
228
- The model has not been aligned for safety. Developers and users should perform their own safety
229
- 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.
230
-
231
-
232
- ### Bias, Risks, and Limitations
233
-
234
- <!-- This section is meant to convey both technical and sociotechnical limitations. -->
235
-
236
- *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.
237
- Like many LLMs, the model can hallucinate and occasionally generates irrelevant content,
238
- introducing fictional elements that are not grounded in the provided context.
239
- Users should also exercise caution in interpreting and validating the model's responses
240
- due to the potential inconsistencies.
241
-
242
- **Limitations**
243
-
244
- In terms of vision capability, Gemma-SEA-LION-v4-27B-IT has been trained and fine-tuned exclusively on the text back-end.
245
- As a result, its vision capabilities are expected to be comparable to those of Gemma 3 IT 27B,
246
- and may not exhibit significant improvements or differences in this area. [🤗 google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it )
247
-
248
-
249
-
250
- ## More Information
251
-
252
- This is the repository for the commercial instruction-tuned model.
253
- The model has not been aligned for safety. Developers and users should perform their own safety
254
- fine-tuning and related security measures. In no event shall the authors be held liable
255
- for any claims, damages, or other liabilities arising from the use of the released weights and codes.
256
-
257
- AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore.
258
- 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.
259
-
260
- For more info, please contact us at [SEA-LION Inquiry Form](https://forms.gle/sLCUVb95wmGf43hi6) or sealion@aisingapore.org
261
-
 
12
  - vi
13
  ---
14