Overview

hishab/titulm-gemma-2-2b-v1.1 is a 2.6 billion parameter language model based on the Gemma 2 architecture, developed by Hishab. It has undergone continual pre-training specifically on a large, curated corpus of Bangla text, totaling 4.4 billion tokens. This specialized training aims to significantly enhance its performance in Bangla language tasks, including text generation and understanding.

Key Capabilities

• Enhanced Bangla Text Generation: Optimized to produce high-quality, fluent Bangla text.
• Superior Bangla Language Understanding: Demonstrates improved performance on various Bangla evaluation benchmarks.
• Monolingual Focus: Primarily designed for Bangla language processing, with expected reduced performance on English tasks compared to the base Gemma 2 model.

Benchmarks and Performance

Evaluation on Bangla benchmark datasets shows that titulm-gemma-2-2b-v1.1 outperforms the base gemma-2-2b model in key areas such as Commonsense QA Bangla, OpenBook QA Bangla, and PIQA Bangla, particularly in 0-shot and 5-shot settings. For instance, it achieves 0.40 in Commonsense QA BN (5-shot) compared to gemma-2-2b's 0.28. While its English capabilities are evaluated, they are intentionally lower than the original gemma-2-2b due to its specialized Bangla training.

Good For

• Bangla text generation applications.
• Bangla language understanding tasks.
• Bangla instruction fine-tuning tasks.