Safetensors
gemma2

additional proofreading

#3
by kiliangoto - opened
Files changed (1) hide show
  1. README.md +54 -142
README.md CHANGED
@@ -8,20 +8,22 @@ language:
8
  - su
9
  license: gemma
10
  ---
11
- # Gemma2 9B CPT Sahabat-AI v1
12
 
13
- **Sahabat-AI** (Indonesian language for “close friends”) is a collection of Large Language Models (LLMs) which has been pretrained and instruct-tuned for Indonesian language and its various dialects. Sahabat-AI ecosystem is co-initiated by Indonesian tech and telecommunication companies: GoTo Group and Indosat Ooredoo Hutchison.
 
 
 
14
 
15
- This is the card for the Gemma2 9B CPT Sahabat-AI v1 base model which has undergone continued pre-training from the [Gemma2 9B CPT SEA-Lionv3 base](https://huggingface.co/aisingapore/gemma2-9b-cpt-sea-lionv3-base) model.
16
 
17
  ## Model Details
18
 
19
  ### Model Description
20
 
21
- The continued pre-training data for Gemma2 9B CPT Sahabat-AI v1 base model encompasses approximately 50B tokens.
22
 
23
- - **Co-initiated by:** PT GoTo Gojek Tokopedia Tbk, Indosat Ooredoo Hutchison
24
  - **Developed by:** PT GoTo Gojek Tokopedia Tbk, AI Singapore
 
25
  - **Model type:** Decoder
26
  - **Languages:** English, Indonesian, Javanese, Sundanese
27
  - **License:** [Gemma Community License](https://ai.google.dev/gemma/terms)
@@ -29,146 +31,57 @@ The continued pre-training data for Gemma2 9B CPT Sahabat-AI v1 base model encom
29
  For tokenisation, the model employs the default tokenizer used in Gemma-2-9B. The model has a context length of 8192.
30
 
31
  ### Benchmark Performance
32
- We evaluated Gemma2 9B CPT Sahabat-AI v1 base model on general language capabilities.
33
 
34
  #### General Language Capabilities
35
  For the evaluation of general language capabilities, we employed the
36
  - [SEA HELM (also known as BHASA) evaluation benchmark](https://arxiv.org/abs/2309.06085v2) across a variety of tasks.
37
  - These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), Abstractive Summarization (Summ), Causal Reasoning (Causal) and Natural Language Inference (NLI).
38
  - We also added support for Javanese and Sundanese for the BHASA tasks whenever applicable
39
- - and the common English tasks from the [HuggingFace LLM Leaderboard](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard).
40
- - These tasks consist of [IFEval, BBH, Math Lvl 5, GPQA, MuSR, and MMLU-PRO.](https://huggingface.co/docs/leaderboards/open_llm_leaderboard/about)
41
- - **Caveat**: Our results differ from the HuggingFace LLM Leaderboard because we have used [VLLM](https://docs.vllm.ai/en/latest/) as our inference platform. VLLM caps the context size at **4096 tokens** while HuggingFace was set to **8192 tokens**.
42
 
43
  Note: SEA HELM is implemented using prompts to elicit answers in a strict format. For all tasks, the model is expected to provide an answer tag from which the answer is automatically extracted. For tasks where options are provided, the answer should comprise one of the pre-defined options. The scores for each task is normalised to account for baseline performance due to random chance.
44
 
45
  The evaluation was done **five-shot** with native prompts on a sample of 100-1000 instances for each dataset.
46
 
47
- #### Results
48
-
49
- #### SEA HELM (also known as BHASA)
50
- <table style="border-collapse: collapse; width: 100%; font-size: 10px">
51
- <tr>
52
- <th style="border: 2px solid black; padding: 8px; font-weight: bold;">Language / Model Name [Base]</th>
53
- <th style="border: 1px solid gray; padding: 8px;">Qwen2-7B</th>
54
- <th style="border: 1px solid gray; padding: 8px;">Qwen2.5-7B</th>
55
- <th style="border: 1px solid gray; padding: 8px;">Llama-3-8B</th>
56
- <th style="border: 1px solid gray; padding: 8px;">Llama-3.1-8B</th>
57
- <th style="border: 1px solid gray; padding: 8px;">sea-lionv2.1-8B</th>
58
- <th style="border: 1px solid gray; padding: 8px;">gemma-2-9B</th>
59
- <th style="border: 1px solid gray; padding: 8px;">sea-lionv3-9B</th>
60
- <th style="border: 1px solid gray; padding: 8px;">sahabatai-v1-8B</th>
61
- <th style="border: 2px solid black; padding: 8px;">sahabatai-v1-9B</th>
62
- </tr>
63
- <tr>
64
- <td style="border: 2px solid black; padding: 8px; font-weight: bold;">Overall (Bahasa Indonesia + Javanese + Sundanese)</td>
65
- <td style="border: 1px solid gray; padding: 8px;">42.776</td>
66
- <td style="border: 1px solid gray; padding: 8px;">46.245</td>
67
- <td style="border: 1px solid gray; padding: 8px;">49.160</td>
68
- <td style="border: 1px solid gray; padding: 8px;">49.577</td>
69
- <td style="border: 1px solid gray; padding: 8px;">48.602</td>
70
- <td style="border: 1px solid gray; padding: 8px;">58.972</td>
71
- <td style="border: 1px solid gray; padding: 8px;">60.913</td>
72
- <td style="border: 1px solid gray; padding: 8px;">59.437</td>
73
- <td style="border: 2px solid black; padding: 8px; background-color: lightgreen;">64.123</td>
74
- </tr>
75
- <tr>
76
- <td style="border: 2px solid black; padding: 8px; font-weight: bold;">Bahasa Indonesia</td>
77
- <td style="border: 1px solid gray; padding: 8px;">49.341</td>
78
- <td style="border: 1px solid gray; padding: 8px;">55.913</td>
79
- <td style="border: 1px solid gray; padding: 8px;">47.865</td>
80
- <td style="border: 1px solid gray; padding: 8px;">48.110</td>
81
- <td style="border: 1px solid gray; padding: 8px;">49.154</td>
82
- <td style="border: 1px solid gray; padding: 8px;">58.572</td>
83
- <td style="border: 1px solid gray; padding: 8px; background-color: lightgreen;">62.437</td>
84
- <td style="border: 1px solid gray; padding: 8px;">53.454</td>
85
- <td style="border: 2px solid black; padding: 8px;">60.040</td>
86
- </tr>
87
- <tr>
88
- <td style="border: 2px solid black; padding: 8px; font-weight: bold;">Javanese</td>
89
- <td style="border: 1px solid gray; padding: 8px;">42.774</td>
90
- <td style="border: 1px solid gray; padding: 8px;">45.917</td>
91
- <td style="border: 1px solid gray; padding: 8px;">54.627</td>
92
- <td style="border: 1px solid gray; padding: 8px;">55.215</td>
93
- <td style="border: 1px solid gray; padding: 8px;">52.728</td>
94
- <td style="border: 1px solid gray; padding: 8px;">63.760</td>
95
- <td style="border: 1px solid gray; padding: 8px;">63.363</td>
96
- <td style="border: 1px solid gray; padding: 8px;">65.048</td>
97
- <td style="border: 2px solid black; padding: 8px; background-color: lightgreen;">69.882</td>
98
- </tr>
99
- <tr>
100
- <td style="border: 2px solid black; padding: 8px; font-weight: bold;">Sundanese</td>
101
- <td style="border: 1px solid gray; padding: 8px;">36.213</td>
102
- <td style="border: 1px solid gray; padding: 8px;">36.905</td>
103
- <td style="border: 1px solid gray; padding: 8px;">44.988</td>
104
- <td style="border: 1px solid gray; padding: 8px;">45.407</td>
105
- <td style="border: 1px solid gray; padding: 8px;">43.925</td>
106
- <td style="border: 1px solid gray; padding: 8px;">54.583</td>
107
- <td style="border: 1px solid gray; padding: 8px;">56.939</td>
108
- <td style="border: 1px solid gray; padding: 8px;">59.809</td>
109
- <td style="border: 2px solid black; padding: 8px; background-color: lightgreen;">62.446</td>
110
- </tr>
111
- </table>
112
-
113
- #### English Results
114
- <table style="border-collapse: collapse; width: 100%; font-size: 10px">
115
- <tr>
116
- <th style="border: 1px solid gray; padding: 8px;">Model Name [BASE]</th>
117
- <th style="border: 1px solid gray; padding: 8px;">Qwen2-7B</th>
118
- <th style="border: 1px solid gray; padding: 8px;">Qwen2.5-7B</th>
119
- <th style="border: 1px solid gray; padding: 8px;">Llama-3-8B</th>
120
- <th style="border: 1px solid gray; padding: 8px;">Llama-3.1-8B</th>
121
- <th style="border: 1px solid gray; padding: 8px;">sea-lionv2.1-8B</th>
122
- <th style="border: 1px solid gray; padding: 8px;">gemma-2-9B</th>
123
- <th style="border: 1px solid gray; padding: 8px;">sea-lionv3-9B</th>
124
- <th style="border: 1px solid gray; padding: 8px;">sahabatai-v1-8B</th>
125
- <th style="border: 2px solid black; padding: 8px;">sahabatai-v1-9B</th>
126
- </tr>
127
- <tr>
128
- <td style="border: 1px solid gray; padding: 8px; font-weight: bold;">Average</td>
129
- <td style="border: 1px solid gray; padding: 8px;">23.68</td>
130
- <td style="border: 1px solid gray; padding: 8px; background-color: lightgreen;">24.65</td>
131
- <td style="border: 1px solid gray; padding: 8px;">13.56</td>
132
- <td style="border: 1px solid gray; padding: 8px;">13.69</td>
133
- <td style="border: 1px solid gray; padding: 8px;">12.77</td>
134
- <td style="border: 1px solid gray; padding: 8px;">13.34</td>
135
- <td style="border: 1px solid gray; padding: 8px;">21.99</td>
136
- <td style="border: 1px solid gray; padding: 8px;">13.92</td>
137
- <td style="border: 2px solid black; padding: 8px;">19.62</td>
138
- </tr>
139
- </table>
140
-
141
 
142
  ## Training Details
143
 
144
  ### Data
145
 
146
- Gemma2 9B CPT Sahabat-AI v1 base model was continued pre-trained on 50B tokens of the following data:
147
 
148
  | Data Source | Unique Tokens (B) | Multiplier | Total Tokens (B) | Percentage (%)|
149
  |---------------------------------------|:-----------------:|:----------:|:----------------:|:-------------:|
150
- | Dolma Refined Web | 9.5 | 1 | 9.5 | 18.7 |
151
- | Dolma arXiv | 0.6 | 1 | 0.6 | 1.18 |
152
- | Stack V2 | 5.5 | 1 | 5.5 | 10.85 |
153
- | Dolma Semantic Scholar | 1.2 | 1 | 1.2 | 2.37 |
154
- | Dolma Reddit | 1.7 | 1 | 1.7 | 3.36 |
155
- | Dolma Pes2o | 1.2 | 1 | 1.2 | 2.37 |
156
- | Wiki* + News* - Indonesian | 1.0 | 1 | 1.0 | 1.97 |
157
- | SEA-LION Pile - Indonesian | 27.0 | 1 | 27.0 | 53.3 |
158
- | JV Pile - Javanese | 0.92 | 1.6 | 1.5 | 3.0 |
159
- | SU Pile - Sundanese | 0.39 | 3.8 | 1.5 | 3.0 |
 
 
 
 
 
160
 
161
  Note:
162
  - All token counts are counted using Gemma2 tokenizer
163
  - Wiki* sources includes Wikipedia, Wiki Books, Wiki Source, Wiki Voyage and Fandom Wiki
164
- - News* sources includes VOA, Global Voices
165
 
166
  ### Infrastructure
167
 
168
- Gemma2 9B CPT Sahabat-AI v1 was trained using [MosaicML Composer](https://github.com/mosaicml/composer)
169
  on the following hardware:
170
 
171
- | Training Details | Gemma2 9B CPT Sahabat-AI v1|
172
  |----------------------|:--------------------------:|
173
  | Nvidia H100 80GB GPU | 32 |
174
  | Training Duration | 7 days |
@@ -176,7 +89,7 @@ on the following hardware:
176
 
177
  ### Configuration
178
 
179
- | HyperParameter | Gemma2 9B CPT Sahabat-AI v1|
180
  |-------------------|:--------------------------:|
181
  | Precision | bfloat16 |
182
  | Optimizer | decoupled_adamw |
@@ -185,30 +98,12 @@ on the following hardware:
185
  | Global Batch Size | 256 |
186
  | Micro Batch Size | 1 |
187
 
188
- ## Call for Collaboration
189
-
190
- Sahabat-AI (Indonesian language for “close friends”) a **local open source Large Language Model (LLM) ecosystem in Indonesian language**, co-initiated by Indonesian tech and telecommunication companies: GoTo Group and Indosat Ooredoo Hutchison.
191
- Sahabat-AI ecosystem aims to empower Indonesians who want to develop AI-based services and applications using Bahasa Indonesia and its various local dialects.
192
-
193
- We are supported by research centers and global tech experts such as AI Singapore and Tech Mahendra to train the model to gain general language understanding.
194
-
195
- We also collaborate with key top Indonesia universities such as University of Indonesia, Gadjah Mada University, Bogor Institute of Agriculture, Bandung Institute of Technology, including top Indonesia media groups, such as Kompas Gramedia Group and Republika to train and enrich the model in Bahasa Indonesia, ensuring optimum provision of local context and cultural relevance.
196
 
197
- We would like to invite **researchers, developers, and language enthusiasts** to actively contribute to the enhancement and expansion of Sahabat-AI.
198
- Your collaborations can involve:
199
- - Identifying and reporting technical issues
200
- - Sharing pre-training, instruction, and preference data
201
- - Improving documentation usability
202
- - Proposing and implementing new model evaluation tasks and metrics
203
-
204
- Join us in shaping the future of Sahabat-AI by sharing your expertise and insights to make these models more accessible, accurate, and versatile.
205
-
206
- You can contribute your ideas through [this form.](https://docs.google.com/forms/d/1_us969eQtEooYOn4XkvGkdP5VHOyCbO6L_sd9kTMnaA/edit)
207
-
208
- ## The Development Team (in ascending alphabetical order)
209
 
210
  ### AI Singapore
211
  Chan Adwin<br>
 
212
  Cheng Nicholas<br>
213
  Choa Esther<br>
214
  Huang Yuli<br>
@@ -239,7 +134,6 @@ Yong Xianbin<br>
239
 
240
  ### PT GoTo Gojek Tokopedia Tbk
241
  Anissa Dininta<br>
242
- Chau Shiau Ching<br>
243
  Choiri Hendra Hadhil<br>
244
  Goel Priyank<br>
245
  Saini Ajay Kumar<br>
@@ -249,15 +143,16 @@ Tep Kilian Rithi<br>
249
  Tiwari Anupam<br>
250
  Widjojo Daniel<br>
251
 
 
252
  ## Acknowledgements
253
 
254
  AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore.
 
255
 
256
- Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore.
257
 
258
  ## Contact
259
 
260
- For more info, please contact us using this [Sahabat-AI Inquiry Form.](https://docs.google.com/forms/d/1_us969eQtEooYOn4XkvGkdP5VHOyCbO6L_sd9kTMnaA/edit)
261
 
262
 
263
  ## Disclaimer
@@ -265,4 +160,21 @@ For more info, please contact us using this [Sahabat-AI Inquiry Form.](https://d
265
  This is the repository for the base model.
266
  The model has _not_ been aligned for safety.
267
  Developers and users should perform their own safety fine-tuning and related security measures.
268
- In no event shall the authors be held liable for any claim, damages, or other liability arising from the use of the released weights and codes.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8
  - su
9
  license: gemma
10
  ---
11
+ # Gemma2 9B CPT Sahabat AI v1
12
 
13
+ Sahabat AI v1 is a collection of Large Language Models (LLMs) which has been pretrained and instruct-tuned for Indonesian languages.
14
+ This is the card for the Gemma2 9B CPT Sahabat AI v1 base model which has undergone continued pre-training from the base [Gemma2 9B CPT SEA-LIONv3 base](https://huggingface.co/aisingapore/gemma2-9b-cpt-sea-lionv3-base) model.
15
+
16
+ Sahabat is Indonesian for "Close Friends."
17
 
 
18
 
19
  ## Model Details
20
 
21
  ### Model Description
22
 
23
+ The continued pre-training data for Gemma2 9B CPT Sahabat AI v1 base model encompasses approximately 50B tokens.
24
 
 
25
  - **Developed by:** PT GoTo Gojek Tokopedia Tbk, AI Singapore
26
+ - **Funded by:** PT GoTo Gojek Tokopedia Tbk, AI Singapore
27
  - **Model type:** Decoder
28
  - **Languages:** English, Indonesian, Javanese, Sundanese
29
  - **License:** [Gemma Community License](https://ai.google.dev/gemma/terms)
 
31
  For tokenisation, the model employs the default tokenizer used in Gemma-2-9B. The model has a context length of 8192.
32
 
33
  ### Benchmark Performance
34
+ We evaluated Gemma2 9B CPT Sahabat AI v1 base model on general language capabilities.
35
 
36
  #### General Language Capabilities
37
  For the evaluation of general language capabilities, we employed the
38
  - [SEA HELM (also known as BHASA) evaluation benchmark](https://arxiv.org/abs/2309.06085v2) across a variety of tasks.
39
  - These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), Abstractive Summarization (Summ), Causal Reasoning (Causal) and Natural Language Inference (NLI).
40
  - We also added support for Javanese and Sundanese for the BHASA tasks whenever applicable
41
+ - [IndoMMLU](https://arxiv.org/pdf/2310.04928)
42
+ - These tasks include examination questions on Humanities, Indonesian language, Local languages and cultures, Social science and STEM across primary, middle, and high school levels.
43
+ - and the well known [English MMLU](https://arxiv.org/pdf/2009.03300)
44
 
45
  Note: SEA HELM is implemented using prompts to elicit answers in a strict format. For all tasks, the model is expected to provide an answer tag from which the answer is automatically extracted. For tasks where options are provided, the answer should comprise one of the pre-defined options. The scores for each task is normalised to account for baseline performance due to random chance.
46
 
47
  The evaluation was done **five-shot** with native prompts on a sample of 100-1000 instances for each dataset.
48
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49
 
50
  ## Training Details
51
 
52
  ### Data
53
 
54
+ Gemma2 9B CPT Sahabat AI v1 base model was continued pre-trained on 50B tokens of the following data:
55
 
56
  | Data Source | Unique Tokens (B) | Multiplier | Total Tokens (B) | Percentage (%)|
57
  |---------------------------------------|:-----------------:|:----------:|:----------------:|:-------------:|
58
+ | Dolma Refined Web | 9.5 | 1 | 9.5 | 17.36 |
59
+ | Dolma arXiv | 0.6 | 1 | 0.6 | 1.10 |
60
+ | Dolma Star Coder | 5.5 | 1 | 5.5 | 10.05 |
61
+ | Dolma Semantic Scholar | 1.2 | 1 | 1.2 | 2.19 |
62
+ | Dolma Reddit | 1.7 | 1 | 1.7 | 3.11 |
63
+ | Dolma C4 | 1.5 | 1 | 1.4 | 2.56 |
64
+ | Wiki* + News* - Indonesian | 1.3 | 4 | 5.2 | 9.50 |
65
+ | SEA-LION Pile - Indonesian | 27.0 | 1 | 27.0 | 49.34 |
66
+ | SEA-LION Pile - Javanese | 0.5 | 1.5 | 0.75 | 1.37 |
67
+ | CC 100 - Javanese | 0.05 | 1.5 | 0.075 | 0.14 |
68
+ | HPLT - Javanese | 0.3 | 1.5 | 0.45 | 0.82 |
69
+ | SEA-LION Pile - Sundanese | 0.2 | 3.6 | 0.75 | 1.37 |
70
+ | CC 100 - Sundanese | 0.02 | 3.6 | 0.075 | 0.14 |
71
+ | HPLT - Sundanese | 0.16 | 3.6 | 0.45 | 0.82 |
72
+ | Others (Javanese, Sundanese) | 0.034 | 2.2 | 0.076 | 0.14 |
73
 
74
  Note:
75
  - All token counts are counted using Gemma2 tokenizer
76
  - Wiki* sources includes Wikipedia, Wiki Books, Wiki Source, Wiki Voyage and Fandom Wiki
77
+ - News* sources includes VOA, Global Voices, MediaCorp
78
 
79
  ### Infrastructure
80
 
81
+ Gemma2 9B CPT Sahabat AI v1 was trained using [MosaicML Composer](https://github.com/mosaicml/composer)
82
  on the following hardware:
83
 
84
+ | Training Details | Gemma2 9B CPT Sahabat AI v1|
85
  |----------------------|:--------------------------:|
86
  | Nvidia H100 80GB GPU | 32 |
87
  | Training Duration | 7 days |
 
89
 
90
  ### Configuration
91
 
92
+ | HyperParameter | Gemma2 9B CPT Sahabat AI v1|
93
  |-------------------|:--------------------------:|
94
  | Precision | bfloat16 |
95
  | Optimizer | decoupled_adamw |
 
98
  | Global Batch Size | 256 |
99
  | Micro Batch Size | 1 |
100
 
 
 
 
 
 
 
 
 
101
 
102
+ ## The Team (by ascending alphabetical order)
 
 
 
 
 
 
 
 
 
 
 
103
 
104
  ### AI Singapore
105
  Chan Adwin<br>
106
+ Chau Shiau Ching<br>
107
  Cheng Nicholas<br>
108
  Choa Esther<br>
109
  Huang Yuli<br>
 
134
 
135
  ### PT GoTo Gojek Tokopedia Tbk
136
  Anissa Dininta<br>
 
137
  Choiri Hendra Hadhil<br>
138
  Goel Priyank<br>
139
  Saini Ajay Kumar<br>
 
143
  Tiwari Anupam<br>
144
  Widjojo Daniel<br>
145
 
146
+ <!--
147
  ## Acknowledgements
148
 
149
  AI Singapore is a national programme supported by the National Research Foundation, Singapore and hosted by the National University of Singapore.
150
+ Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore. -->
151
 
 
152
 
153
  ## Contact
154
 
155
+ For more info, please contact us using this [Sahabat Inquiry Form.](https://docs.google.com/forms/d/1_us969eQtEooYOn4XkvGkdP5VHOyCbO6L_sd9kTMnaA/edit)
156
 
157
 
158
  ## Disclaimer
 
160
  This is the repository for the base model.
161
  The model has _not_ been aligned for safety.
162
  Developers and users should perform their own safety fine-tuning and related security measures.
163
+ In no event shall the authors be held liable for any claim, damages, or other liability arising from the use of the released weights and codes.
164
+
165
+
166
+ ## References
167
+ ### IndoMMLU Reference
168
+
169
+ ```bibtex
170
+ @inproceedings{koto-etal-2023-indommlu,
171
+ title = "Large Language Models Only Pass Primary School Exams in {I}ndonesia: A Comprehensive Test on {I}ndo{MMLU}",
172
+ author = "Fajri Koto and Nurul Aisyah and Haonan Li and Timothy Baldwin",
173
+ booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
174
+ month = December,
175
+ year = "2023",
176
+ address = "Singapore",
177
+ publisher = "Association for Computational Linguistics",
178
+ }
179
+ }
180
+ ```