alphaedge-ai/gemma-3-270m-it-fas-16384
The alphaedge-ai/gemma-3-270m-it-fas-16384 is a 0.3 billion parameter instruction-tuned causal language model, derived from Google's Gemma-3-270m-it. It features a significantly reduced vocabulary size of 16,384 tokens, resulting in a 58.67% smaller model size compared to its base. This optimization specifically targets the Persian language, making it highly efficient for Persian-centric applications while maintaining similar performance to the original model.
Loading preview...
Overview
This model, alphaedge-ai/gemma-3-270m-it-fas-16384, is a specialized version of the google/gemma-3-270m-it model, optimized for the Persian language. It achieves a 58.67% reduction in model size by significantly trimming its vocabulary from 262,144 tokens down to just 16,384 tokens. This process, known as "trimming," focuses on removing tokens not commonly used in Persian, leading to a much smaller memory footprint while aiming for comparable performance to the original model in its target language.
Key Capabilities
- Highly optimized for Persian language: Specifically designed for efficient processing and generation in Persian.
- Reduced memory footprint: Achieves a 58.67% smaller model size (110,811,776 parameters) compared to the base model, making it suitable for resource-constrained environments.
- Efficient vocabulary: Features a 93.75% reduction in vocabulary size, from 262,144 to 16,384 tokens, tailored for Persian.
Good for
- Applications requiring a compact and efficient instruction-tuned model for Persian language tasks.
- Deployments where memory and computational resources are limited, but strong Persian language capabilities are essential.
- Use cases where the performance for other languages is not a primary concern, as the trimming process may impact non-Persian language performance.