GanitLLM-0.6B_SFT_CGRPO: Bengali Mathematical Reasoning Model

GanitLLM-0.6B_SFT_CGRPO is a 0.6 billion parameter causal language model developed by dipta007, built upon the Qwen3-0.6B architecture. It is specifically designed for mathematical reasoning in Bengali, making it ideal for applications requiring strong performance in this domain, especially in resource-constrained environments.

Key Capabilities and Differentiators

• Enhanced Bengali Mathematical Reasoning: Achieves significant improvements in accuracy on Bengali mathematical benchmarks, including a +20.0 accuracy increase on Bn-MGSM (from 8.4 to 28.4) and a +40.2 accuracy increase on Bn-MSVAMP (from 12.2 to 52.4) compared to its base model.
• Efficient and Bengali-Centric Solutions: Demonstrates 88.50% Bengali reasoning in its outputs and generates solutions with 80.4% fewer tokens (averaging 248 words vs. 1265 words for the base model), indicating more concise and relevant responses.
• Advanced Training Methodology: Utilizes a multi-stage training pipeline involving Supervised Fine-Tuning (SFT) on ~11k examples and a novel Curriculum-GRPO (Reinforcement Learning with difficulty-aware sampling) approach on ~7.3k examples. This includes specific reward functions for format validation, correctness (Bengali and English answer match), and ensuring high Bengali text percentage in reasoning.

Use Cases

This model is particularly well-suited for:

• Bengali mathematical problem-solving applications.
• Educational tools requiring step-by-step reasoning in Bengali.
• Deployments where computational resources are limited, due to its small parameter count.

For more details, refer to the project page and the associated arXiv paper.