Hugging Face Overview
GanitLLM-1.7B-SFT: Bengali Mathematical Reasoning Model
GanitLLM-1.7B-SFT is a 1.7 billion parameter causal language model, fine-tuned by dipta007 from the Qwen3-1.7B base model. It specializes in Bengali mathematical reasoning, having undergone Supervised Fine-Tuning (SFT) on the GANIT dataset, which comprises approximately 11,000 Bengali math problems with chain-of-thought reasoning.
Key Capabilities & Performance
- Enhanced Bengali Math Reasoning: Achieves 87.79% Bengali reasoning accuracy, a substantial improvement over the base model's 19.64%.
- Benchmark Improvements: Shows significant gains on Bengali math benchmarks, with a +33.60 accuracy increase on Bn-MGSM (from 15.20 to 48.80) and a +50.50 accuracy increase on Bn-MSVAMP (from 14.10 to 64.60).
- Concise Solutions: Generates solutions with 77.5% fewer words (253 words vs. 1124 words for the base model), indicating more efficient reasoning.
- Multilingual Support: Supports both Bengali and English, with a context length of 4,096 tokens.
Use Cases
This model is ideal for applications requiring accurate and efficient mathematical problem-solving in Bengali. It serves as a strong foundation for further reinforcement learning (RL) training, with more advanced RL-enhanced versions like GanitLLM-1.7B_SFT_CGRPO and GanitLLM-1.7B_SFT_GRPO available for optimal performance.