maanka2/somgpt

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.3BQuant:BF16Context Size:32kPublished:May 25, 2026License:gemmaArchitecture:Transformer Featherless Exclusive Cold

maanka2/somgpt is a 0.3 billion parameter Somali causal language model, continued from Google's Gemma-3-270m architecture. It was further pre-trained on the maanka2/somali-web-corpus to enhance its understanding of Somali vocabulary, grammar, and writing patterns. This base model is designed for text continuation and language modeling in Somali, serving as a foundation for further fine-tuning.

Loading preview...

SOMGPT: A Somali Causal Language Model

SOMGPT is a 0.3 billion parameter causal language model developed by maanka2, built upon the google/gemma-3-270m architecture. This model has undergone extensive pre-training on the maanka2/somali-web-corpus, a comprehensive dataset of cleaned Somali-language web content. The primary goal of this additional pre-training was to significantly improve the model's proficiency in Somali vocabulary, grammar, spelling, and overall writing patterns.

Key Capabilities

  • Somali Language Understanding: Enhanced comprehension and generation of Somali text due to specialized pre-training.
  • Text Continuation: Designed to predict and generate subsequent text based on a given prompt in Somali.
  • Language Modeling: Excels at understanding the statistical structure and patterns of the Somali language.

Intended Use Cases

SOMGPT is a base language model and is specifically optimized for:

  • Foundational Somali NLP Research: Serving as a robust starting point for various Somali natural language processing tasks.
  • Further Fine-tuning: Ideal for developers looking to build specialized applications by applying additional supervised fine-tuning (SFT) or instruction tuning for tasks like:
    • Chatbots
    • Question Answering systems
    • Assistant-style interactions

It is important to note that SOMGPT is not instruction-tuned out-of-the-box and is not designed for direct conversational or assistant-style use without further customization.