gemma3-cpt-gotoai

GoTo-AI is a collection of Large Language Models (LLMs) which have been pretrained and instruct-tuned for Indonesian culture and languages. GoTo-AI-Gemma-3-27B-IT leverages the Gemma-3 architecture alongside specialized local models.

• Initiated by: PT GoTo Gojek Tokopedia Tbk
• Developed by: PT GoTo Gojek Tokopedia Tbk, AI Singapore
• Model type: Decoder
• Languages supported: English, Indonesian, Javanese, Sundanese, Batak Toba, Balinese
• License: Gemma License

Model Details

Model Description

For tokenisation, the model employs the default Gemma 3 tokenizer. The model has a context length of 128k tokens.

Benchmark Performance

We evaluate the models in the GoTo-AI ecosystem on both general language capabilities and instruction-following capabilities.

IndoMMLU

For the evaluation of capabilities rooted in the local Indonesian context, we employed the IndoMMLU evaluation benchmark across a variety of tasks. These tasks include examination questions on Humanities, Indonesian language, Local languages and cultures, Social science and STEM across primary, middle, and high school levels.

Note: IndoMMLU 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.

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).

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.

The evaluation was done zero-shot with native prompts on a sample of 100-1000 instances for each dataset.

Instruction-following Capabilities

Since this is an instruction-following model, we also evaluated it on instruction-following capabilities with two datasets, SEA-IFEval (based on IFEval) and SEA-MTBench (based on MT-Bench).

As these 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.

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).

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). A tie is given a score of 0.5. For more details on benchmark performance across the ecosystem, please refer to the GoTo-AI leaderboard.

Usage

This model can be run using the Transformers library. To load this 27B model in FP16 or BF16 precision, you need a minimum of approximately 54-60 GB of VRAM.

Recommended setups include:

• 1× NVIDIA A100 (80GB)
• 2× NVIDIA L40s
• 3× NVIDIA RTX 4090

import transformers
import torch

# Update model_id to your actual Hugging Face repo path
model_id = "GoToCompany/GoTo-AI-Gemma-3-27B-IT"

pipeline = transformers.pipeline(
   "text-generation",
   model=model_id,
   model_kwargs={"torch_dtype": torch.bfloat16},
   device_map="auto",
)

messages = [
   {"role": "user", "content": "Sopo wae sing ana ing Punakawan?"},
]

outputs = pipeline(
   messages,
   max_new_tokens=256,
)
print(outputs[0]["generated_text"][-1])

Limitations

Disclaimer

This is the repository for the Instruct model. While the model has been aligned for general 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 claim, damages, or other liability arising from the use of the released weights and codes.

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 generate 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 in its reasoning. Please refer to our responsible use guide.

Safety Fine-Tuning

The base models in the GoTo-AI ecosystem have bee