manthilaffs/Gamunu-4B-Instruct-Alpha

VISIONConcurrency Cost:1Model Size:4.3BQuant:BF16Ctx Length:32kPublished:Oct 29, 2025License:apache-2.0Architecture:Transformer Open Weights Cold

Gamunu-4B-Instruct-Alpha by Manthila Mallawa is a 4.3 billion parameter experimental Sinhala-centric bilingual instruction-tuned language model, built on Google's Gemma 3 4B. It excels in fluent, idiomatic Sinhala generation, robust Sinhala-English understanding, and mathematical reasoning. This model is primarily designed for research, benchmarking, and controlled deployments in educational, cultural, and academic content generation.

Loading preview...

Gamunu-4B-Instruct-Alpha: A Sinhala-Centric Bilingual LLM

Gamunu-4B-Instruct-Alpha is an experimental 4.3 billion parameter language model developed by Manthila Mallawa as part of The Gamunu Project. It is the first checkpoint in a series of Sinhala-centric bilingual LLMs, built upon Google's Gemma 3 4B base model. The model underwent continued pre-training to enhance Sinhala linguistic coverage and then supervised fine-tuning on a custom Sinhala instruction dataset, focusing on reasoning, roleplay, and assistant-style behavior.

Key Capabilities

  • Bilingual Fluency: Generates fluent, idiomatic Sinhala and demonstrates robust Sinhala ↔ English understanding.
  • Reasoning: Exhibits solid mathematical reasoning (percentages, word problems) and logical, step-by-step reasoning in QA tasks.
  • Instruction Following: Accurately adheres to single-turn instructions and can simulate expert personas (teacher, scientist, analyst).
  • NLP Tasks: Supports text generation, summarization, translation (Sinhala ↔ English), paraphrasing, question answering, and instruction-based classification.

Limitations

As an alpha experimental model, it currently lacks conversational memory, is single-turn only, and has not undergone RLHF or safety tuning. Users may encounter occasional factual drift.

Intended Use Cases

  • Research and evaluation of Sinhala LLMs.
  • Educational assistants and analytical Q&A.
  • Cultural, marketing, and academic content generation.
  • Benchmarking instruction following in low-resource languages.