Padauk

Overview

Padauk explores the assistant layer of the Burmese GPT project family: understanding Burmese requests, structuring tasks, and deciding when external tools can help. The work focuses on practical uses such as writing, summarization, search, and coding assistance.

The project documentation describes a Gemma 4 base adapted with a custom xLAM-format dataset for Burmese intent understanding and agentic behavior. The linked model artifacts provide the release-specific details.

Implementation

The implementation separates direct conversation from a tool-assisted mode. The latter plans tool calls for web search and date/time information, then passes the resulting context to the language model. Responses stream through server-side API handlers.

The deployment work covers GGUF model packaging, Ollama-compatible serving, and OpenAI-compatible interfaces. A document translation prototype also explored extracting English PDF or text content in the browser and translating it in chunks.

Scope and limitations

This is a research and implementation record, not an active hosted assistant service. The conversation and document translation demos are unavailable.

Model output requires review. Search-backed responses depend on source quality, and local inference performance depends on the chosen model artifact and available hardware. The deployment notes describe the configurations used during development, rather than a performance guarantee.

Model artifacts and sources

• Download the Ollama-ready Padauk release
• View the original fine-tuned weights
• BurmeseGPT package reference

Related work and reading

• Padauk and low-resource deployment
• Deploying a fine-tuned language model
• Burmese GPT · Burmese-Coder-4B