Yway: A System One-Style Decision Model for Burmese

I’ve just released Yway (ရွေး, “choose”), an open AI model designed for making structured decisions from Burmese text.

🤗 Model: https://huggingface.co/aungthuhein-dev/yway-system1-xlmr
🧪 Try it: https://yway-wiki-explorer-site.vercel.app/#try

Why another Burmese AI model?

Most language models are built around text generation.

You give them a prompt, and they generate tokens.

For many applications, however, we don't actually need a model to generate text. We need it to answer a specific question, choose between options, or estimate whether something is true.

This is what interested me about the System One approach behind Jev and Laya.

Instead of treating every task as free-form text generation, the model can receive a typed question and produce a structured decision with a probability.

I wanted to explore what this approach could look like for a low-resource language like Burmese.

That led to Yway.

What is Yway?

Yway is a Burmese-focused decision model built on XLM-RoBERTa.

The name ရွေး (Yway) means “choose” in Burmese.

The basic idea is:

Burmese text/context
        +
Typed question
        +
Possible decisions
        ↓
      Yway
        ↓
Decision + probability

Instead of generating a paragraph as an answer, the model makes a structured prediction.

For example:

Context: A Burmese news passage
Question: What is this passage mainly about?
Options: Politics / Economy / Technology / Sports

Yway can select an option and provide its probability.

What can it do?

The current model supports several types of decision tasks, including:

• Topic classification
• Passage relevance
• Question-answer matching
• Answer correctness
• Yes/No decisions
• Choice decisions
• Ordered-score decisions

This makes the model useful for applications where the output needs to be structured rather than free-form.

Why this matters for Burmese

Burmese is still a relatively low-resource language compared with languages such as English.

There are many challenges around Burmese NLP, including limited datasets, limited evaluation resources, and fewer specialized models.

A lot of current work focuses on adapting general-purpose multilingual language models.

Yway explores a different question:

Can decision-model architectures be useful for Burmese without requiring a large generative language model?

This is an early experiment toward answering that question.

Evaluation

On the current test set, Yway achieved the following results:

These numbers are from the current version and should not be interpreted as a comprehensive benchmark of Burmese NLP.

There is still a lot of room for improvement, especially in dataset quality, calibration, and generalization.

Try it yourself

I also built a small interactive demo so you don't need to download the model to experiment with it.

🧪 Yway Wiki Explorer:
https://yway-wiki-explorer-site.vercel.app/#try

You can enter Burmese text and questions and see how the model makes its decision.

What's next?

I’m particularly interested in exploring Yway as a building block for Burmese applications that need fast, structured decisions rather than open-ended generation.

Some possible applications include:

• Burmese information retrieval
• Question-answer verification
• Content classification
• Educational systems
• Document filtering
• RAG pipelines
• Data quality checking
• Lightweight AI agents

I also want to explore whether similar approaches can work for other low-resource languages.

An early experiment

I’m still treating Yway as an experiment rather than a finished production model.

I’m interested in feedback on three things:

1. The model: Where does it fail?
2. The architecture: Are decision models useful for Burmese applications?
3. The use cases: What would you build with a Burmese decision model?

If you work on Burmese NLP, low-resource languages, or decision models, I’d especially love to hear your thoughts.

🤗 Yway on Hugging Face:
https://huggingface.co/aungthuhein-dev/yway-system1-xlmr

🧪 Interactive demo:
https://yway-wiki-explorer-site.vercel.app/#try

Yway is an independent project inspired by the System One model paradigm behind Jev and Laya. It is not affiliated with TypeSafe or ConvAI Innovations.