Hugging Face Overview
GanitLLM-4B-SFT: Bengali Mathematical Reasoning Model
GanitLLM-4B-SFT is a 4 billion parameter causal language model, built upon the Qwen3-4B architecture, and specifically optimized for mathematical reasoning in Bengali. It was developed by dipta007 through Supervised Fine-Tuning (SFT) on the GANIT dataset, which comprises 11,023 Bengali math problems with chain-of-thought reasoning.
Key Capabilities and Performance
This model demonstrates substantial improvements over its base model in Bengali mathematical tasks:
- Enhanced Accuracy: Achieves a +4.80 accuracy increase on the Bn-MGSM benchmark (from 69.20 to 74.00) and a +4.10 accuracy increase on the Bn-MSVAMP benchmark (from 70.50 to 74.60).
- Superior Bengali Reasoning: Exhibits 86.65% Bengali reasoning capability, a significant leap from the base model's 14.79%.
- Concise Solutions: Generates solutions with 80.5% fewer words (184 words vs. 943 words for the base model), making outputs more efficient.
- Multilingual Support: Supports both Bengali and English, with a context length of 4,096 tokens.
Intended Use Cases
GanitLLM-4B-SFT is ideal for applications requiring:
- Bengali Mathematical Problem Solving: Excels at providing step-by-step reasoning for math problems in Bengali.
- Foundation for Further RL Training: Serves as the SFT-only checkpoint, designed to be a strong base for subsequent Reinforcement Learning (RL) enhancements (e.g., GRPO/CGRPO versions like GanitLLM-4B_SFT_CGRPO).
This model is a specialized tool for developers focusing on educational technology, localized AI assistants, or any system requiring robust mathematical reasoning in the Bengali language.