Update README

README.md

@@ -21,18 +21,18 @@ license: gemma
 base_model_relation: finetune
 
 ---
-*Gemma-SEA-LION-v4-27B (Base Model) Last updated: 2025-08-18*
+*Gemma-SEA-LION-v4-27B-IT (IT Model) Last updated: 2025-08-18*
 
 ---
 
-# Model Card for Gemma-SEA-LION-v4-27B
+# Model Card for Gemma-SEA-LION-v4-27B-IT
 
 <!-- Provide a quick summary of what the model is/does. -->
 
 **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 is a multilingual model which has undergone continued pre-training on 
+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.
 
@@ -46,7 +46,7 @@ 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.
+a decoder model using the Gemma 3 architecture, to create Gemma-SEA-LION-v4-27B-IT.
 
 For tokenization, the model employs the default tokenizer used in Gemma 3 27B IT.
 
@@ -58,7 +58,7 @@ For tokenization, the model employs the default tokenizer used in Gemma 3 27B IT
 - **Context length:** 128k
 - **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-3-27B-IT](https://huggingface.co/google/gemma-3-27b-it)
+- **Finetuned from model:** Gemma-SEA-LION-v4-27B
 
 ### Model Sources
 
@@ -92,7 +92,7 @@ due to the potential inconsistencies.
 
 **Limitations**
 
-In terms of vision capability, Gemma-SEA-LION-v4-27B has been trained and fine-tuned exclusively on the text back-end.
+In terms of vision capability, Gemma-SEA-LION-v4-27B-IT has been trained and fine-tuned exclusively on the text back-end.
 As a result, its vision capabilities are expected to be comparable to those of Gemma 3 IT 27B, 
 and may not exhibit significant improvements or differences in this area. [🤗 google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it )
 
@@ -110,7 +110,7 @@ import torch
 
 pipe = pipeline(
     "text-generation",
-    model="aisingapore/Gemma-SEA-LION-v4-27B",
+    model="aisingapore/Gemma-SEA-LION-v4-27B-IT",
     device="cuda",
     torch_dtype=torch.bfloat16
 )
@@ -143,47 +143,17 @@ 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.
 
-The 500 billion tokens are sampled from a much larger pool of 1 trillion tokens from open-sourced datasets with the optimal datamix shown below determined by our experiments.
-
-
-| Language                         | Dataset Name             | Total Tokens (B) | Percentage (%) | Total percentage (%) |
-|-----------------------------------|-------------------------|------------------|----------------|---------------------|
-| Code                              | StarCoder (OLMo 2 Version) | 50B             | 10             | 10                  |
-| EN                                | Fineweb-Edu             | 80B              | 16             | 40                  |
-|                                  | DCLM-OLMo2-HQ           | 80B              | 16             |                     |
-|                                  | Non-CC-EN               | 40B              | 8              |                     |
-| ZH                                | SEA-LION Pile v1        | 13.5B            | 2.7            | 9                   |
-|                                  | Fineweb2                | 13.5B            | 2.7            |                     |
-|                                  | Fineweb2-HQ             | 4.5B             | 0.9            |                     |
-| VI                                | SEA-LION Pile v1        | 4.25B            | 0.85           | 8.5                 |
-|                                  | SEA-LION Pile v2        | 12.75B           | 2.55           |                     |
-|                                  | Fineweb2                | 8.5B             | 1.7            |                     |
-|                                  | Non-CC-VI               | 17B              | 3.4            |                     |
-| ID                                | SEA-LION Pile v1        | 5.66B            | 1.13           | 8.5                 |
-|                                  | SEA-LION Pile v2        | 17B              | 3.4            |                     |
-|                                  | Fineweb2                | 11.33B           | 2.27           |                     |
-|                                  | Non-CC-ID               | 8.5B             | 1.7            |                     |
-| TH                                | SEA-LION Pile v1        | 3.035B           | 0.61           | 8.5                 |
-|                                  | SEA-LION Pile v2        | 9.107B           | 1.82           |                     |
-|                                  | Fineweb2                | 3.035B           | 0.61           |                     |
-|                                  | WangChanBERTa           | 3.035B           | 0.61           |                     |
-|                                  | Dolmav1                 | 3.035B           | 0.61           |                     |
-|                                  | Non-CC-TH               | 21.25B           | 4.25           |                     |
-| TL, TA, MS, KM, LO and MY         | ALL_LANG                | 77.5B            | 15.5           | 15.5                |
-
-
-
 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). 
+- SEA-LION 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.
+- SEA-LION 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/)
 
@@ -194,16 +164,17 @@ The cutoff date of this version is September 2020.
 
 #### Training Hyperparameters
 
-- **Training regime:** 
-
-
-| Hyperparameter    | Gemma-SEA-LION-v4-27B |
-|-------------------|-----------------------|
-| Precision         | bfloat16              |
-| Optimizer         | decoupled_adamw       |
-| Scheduler         | CosineAnnealing       |
-| Learning Rate     | 4.00E-08              |
-| Global Batch Size | 1024                  |
+- **Training regime:** We perform post-training using a variety of Reinforcement Learning (RL) methods. 
+The instruction fine-tuning dataset combines our SEA-Instruct, Infinity-Instruct, 
+and OpenMath-Instruct 2 with open-source datasets such as 
+nvidia/Llama-Nemotron-Post-Training-Dataset (RL set) and zwhe99/DeepMath-103K. 
+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 
+the *rejected* outputs.
  <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
 
 
@@ -217,11 +188,11 @@ The cutoff date of this version is September 2020.
 
 <!-- This should link to a Dataset Card if possible. -->
 
-We evaluated Gemma-SEA-LION-v4-27B on general language capabilities.
+We evaluated Gemma-SEA-LION-v4-27B-IT on both general language capabilities and instruction-following capabilities.
 
 **Testing Data**
 
-General NLP Behaviour
+General 
 
 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), 
@@ -229,26 +200,47 @@ Toxicity Detection (Toxicity), Translation in both directions (Eng>Lang & Lang>E
 Abstractive Summarisation (Abssum), Causal Reasoning (Causal), Natural Language Inference (NLI), 
 and linguistic diagnostics (LINDSEA).
 
+Instruction-following
+
+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.
+
 
 #### Factors
 
 <!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
 
-Our evaluations were set based on task. 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.
+For instruction-following tasks, our evaluations were organised based on each specific task.
+
+SEA-IFEval (more languages)
+
+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).
+
+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).
 
 
 #### Metrics
 
 <!-- These are the evaluation metrics being used, ideally with a description of why. -->
 
-The evaluation was done **five-shot** with native prompts on a sample of 100-1000 instances for each dataset.
+The evaluation was done **zero-shot** with native prompts on a sample of 100-1000 instances for each dataset.
 
 ### Results
 
-For details on Gemma-SEA-LION-v4-27B performance, please refer to the SEA-HELM leaderboard, [Leaderboard results on SEA-HELM](https://leaderboard.sea-lion.ai/).
+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/).
 
 
 #### Summary