Update README

README.md

@@ -21,22 +21,19 @@ license: gemma
 base_model_relation: finetune
 
 ---
-*Gemma-SEA-LION-v4-27B-IT (IT Model) Last updated: 2025-08-18*
+*Gemma-SEA-LION-v4-27B-IT (IT Model) Last updated: 2025-08-20*
 
 ---
 
 # Model Card for Gemma-SEA-LION-v4-27B-IT
 
 <!-- Provide a quick summary of what the model is/does. -->
-Last updated: 2025-08-19
+
+Last updated: 2025-08-20
 
 **SEA-LION** is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned 
 for the Southeast Asia (SEA) region.
 
-Gemma-SEA-LION-v4-27B-IT is a multilingual model which has undergone continued pre-training on 
-approximately **500B** tokens across 11 SEA languages: Bahasa Indonesia, Burmese, Chinese, English, 
-Khmer, Lao, Malay, Tagalog, Tamil, Thai and Vietnamese.
-
 
 ## Model Details
 
@@ -46,8 +43,13 @@ Khmer, Lao, Malay, Tagalog, Tamil, Thai and Vietnamese.
 
 SEA-LION stands for *Southeast Asian Languages In One Network*. 
 
-We performed continued pre-training in English and SEA languages on Gemma 3 27B IT, 
-a decoder model using the Gemma 3 architecture, to create Gemma-SEA-LION-v4-27B-IT.
+Gemma-SEA-LION-v4-27B is based on Gemma 3 (which supports over 100 languages) and is a multilingual model 
+which has undergone continued pre-training on approximately **500B** tokens across 11 SEA languages: 
+Bahasa Indonesia, Burmese, Chinese, English, Khmer, Lao, Malay, Tagalog, Tamil, Thai and Vietnamese. 
+
+We continued post-training on Gemma-SEA-LION-v4-27B using a QA pairs dataset in 
+Bahasa Indonesia, Burmese, Chinese, English, Khmer, Lao, Malay, Tagalog, Tamil, Thai, and Vietnamese 
+to create *Gemma-SEA-LION-v4-27B-IT*.
 
 For tokenization, the model employs the default tokenizer used in Gemma 3 27B IT.
 
@@ -56,7 +58,7 @@ For tokenization, the model employs the default tokenizer used in Gemma 3 27B IT
 - **Funded by:** Singapore NRF
 - **Shared by:** Products Pillar, AI Singapore
 - **Model type:** Decoder
-- **Context length:** 128k
+- **Context length:** 128k tokens
 - **Language(s) (NLP):**  Bahasa Indonesia, Burmese, Chinese, English, Khmer, Lao, Malay, Tagalog, Tamil, Thai and Vietnamese
 - **License:** [Gemma Terms of Use](https://ai.google.dev/gemma/terms)
 - **Finetuned from model:** Gemma-SEA-LION-v4-27B
@@ -85,7 +87,7 @@ fine-tuning and related security measures. In no event shall the authors be held
 
 <!-- This section is meant to convey both technical and sociotechnical limitations. -->
 
-It is important for users to be aware that our model exhibits certain limitations that warrant consideration. 
+*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 
@@ -144,20 +146,6 @@ The dataset comprises Bahasa Indonesia, Burmese, Chinese, English, Khmer, Lao, M
 Thai and Vietnamese languages, collected from a mixture of sources including web data, code, open-source datasets, 
 and synthetically generated datasets, amounting to a total of 500 billion tokens.
 
-Note:
-
-- All token counts are counted using Gemma 3 tokenizer.
-
-- Pre-training was conducted with batches of 8k token lengths.
-
-- SEA-Pile v1 is processed from Common Crawl WET, which is published [here](https://huggingface.co/datasets/aisingapore/sea-lion-pile). 
-The main proportion is from mC4 dataset (corpus [link](https://huggingface.co/datasets/bertin-project/mc4-sampling)). 
-The cutoff date of this version is September 2020.
-
-- SEA-Pile v2 is processed from Common Crawl WARC from October 2020 to April 2024.
-
-- Tamil news is sourced with permission from [Seithi](https://seithi.mediacorp.sg/)
-
 
 ### Training Procedure
 
@@ -174,7 +162,7 @@ Prompt sampling is guided by a gradient-based analysis process.
 Our post-training workflow consists of multiple stages: instruction fine-tuning, 
 model merging, online RL for both instruction following and math using DRGPPO, 
 and on-policy alignment via APO. For alignment, rejected-chosen pairs are generated 
-from the target model, with the “chosen” responses obtained by rewriting and improving upon 
+from the target model, with the *chosen* responses obtained by rewriting and improving upon 
 the *rejected* outputs.
  <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
 
@@ -189,58 +177,75 @@ the *rejected* outputs.
 
 <!-- This should link to a Dataset Card if possible. -->
 
-We evaluated Gemma-SEA-LION-v4-27B-IT on both general language capabilities and instruction-following capabilities.
+We evaluated Gemma-SEA-LION-v4-27B-IT on general language, multi-turn chat and instruction-following capabilities.
 
 **Testing Data**
 
-General 
+General language capabilities
 
-For the evaluation of general language capabilities, we employed the SEA-HELM evaluation benchmark
-across a variety of tasks. These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), 
-Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), 
-Abstractive Summarisation (Abssum), Causal Reasoning (Causal), Natural Language Inference (NLI), 
-and linguistic diagnostics (LINDSEA).
+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. 
+These tasks include Question Answering (QA), Sentiment Analysis (Sentiment), Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>Eng), 
+Abstractive Summarisation (Abssum), Causal Reasoning (Causal), Natural Language Inference (NLI), Linguistic Diagnostics (LINDSEA), Cultural Knowledge (Kalahi) 
+and Global MMLU Lite.
 
-Instruction-following
+Instruction-following and Multi-turn Chat
 
-We evaluated the models on instruction-following capabilities with two datasets,
-SEA-IFEval (based on IFEval) and SEA-MTBench (based on MT-Bench). 
-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.
+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. 
+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.
 
 
 #### Factors
 
 <!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
 
-For instruction-following tasks, our evaluations were organised based on each specific task.
+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.
+
+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.
+
+The evaluation was done zero-shot with native prompts on a sample of 100-1000 instances for each dataset. 
 
-SEA-IFEval (more languages)
+SEA-IFEval
 
 SEA-IFEval evaluates a model's ability to adhere to constraints provided in the prompt, 
-for example beginning a response with a specific word/phrase or answering with 
-a certain number of sections. Additionally, accuracy is normalised by the proportion of responses 
-in the correct language (if the model performs the task correctly but responds in the wrong language, 
-it is judged to have failed the task).
+for example beginning a response with a specific word/phrase or answering with a certain number of sections. 
+Additionally, accuracy is normalised by the proportion of responses in the correct language 
+(if the model performs the task correctly but responds in the wrong language, it is judged to have failed the task).
 
 SEA-MTBench
 
-SEA-MTBench evaluates a model's ability to engage in multi-turn (2 turns) conversations and 
-respond in ways that align with human needs. We use gpt-4-1106-preview as the judge model and 
-compare against gpt-3.5-turbo-0125 as the baseline model. The metric used is the weighted win rate 
-against the baseline model (i.e. average win rate across each category: Math, Reasoning, STEM, Humanities, Roleplay, Writing, Extraction).
+SEA-MTBench evaluates a model's ability to engage in multi-turn (2 turns) conversations and respond in ways that align with human needs. 
+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. 
+The metric used is the weighted win rate against the baseline model (i.e. average win rate across each category: 
+Math, Reasoning, STEM, Humanities, Roleplay, Writing, Extraction).
 
 
 #### Metrics
 
 <!-- These are the evaluation metrics being used, ideally with a description of why. -->
 
-The evaluation was done **zero-shot** with native prompts on a sample of 100-1000 instances for each dataset.
+The following metrics were used:
+| Task                                | Metric                                 |
+|--------------------------------------|----------------------------------------|
+| Sentiment Analysis                   | Accuracy                               |
+| Extractive QA (ID, VI, TH, TA)       | ChrF++                                 |
+| MCQ-QA (TL, MY, MS)                  | Accuracy                               |
+| Metaphor                             | Accuracy                               |
+| Abstractive Summarisation            | Rouge-L                                |
+| Translations                         | MetricX-24 score (with reference)      |
+| Causal Reasoning                     | Accuracy                               |
+| Natural Language Inference           | Accuracy                               |
+| LINDSEA                              | Accuracy                               |
+| Global MMLU Lite                     | Accuracy                               |
+| Kalahi                               | Accuracy                               |
+| SEA-IFEval                           | Accuracy                               |
+| SEA-MTBench                          | Win rate against a reference           |
+| Toxicity Detection                   | Accuracy                               |
+
 
 ### Results
 
+Coming soon!
+
 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/).
 
 
@@ -268,6 +273,9 @@ The model has not been aligned for safety. Developers and users should perform t
 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.
 
+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.
+
 For more info, please contact us at sealion@aisingapore.org