dipta007/GanitLLM-4B-SFT
GanitLLM-4B-SFT is a 4 billion parameter causal language model developed by dipta007, based on Qwen3-4B, and fine-tuned for Bengali mathematical reasoning. It significantly improves accuracy on Bengali math benchmarks (Bn-MGSM, Bn-MSVAMP) and generates more concise solutions compared to its base model. This model is specifically optimized for solving mathematical problems in Bengali, serving as a foundation for further reinforcement learning.
Loading preview...
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 fine-tuned using Supervised Fine-Tuning (SFT) on the GANIT dataset. This model is designed to excel in mathematical reasoning tasks in Bengali, offering substantial improvements over its base model.
Key Capabilities and Performance
- Enhanced Bengali Mathematical Reasoning: Achieves 86.65% Bengali reasoning, a significant increase from the base model's 14.79%.
- Improved Accuracy: Demonstrates a +4.80 accuracy gain on the Bn-MGSM benchmark (from 69.20 to 74.00) and a +4.10 accuracy gain on the Bn-MSVAMP benchmark (from 70.50 to 74.60).
- Concise Solutions: Generates solutions with 80.5% fewer words (184 words vs. 943 words for the base model), indicating more efficient and direct reasoning.
- Context Length: Supports a context length of 4,096 tokens.
Training Details
The model underwent a single-stage Supervised Fine-Tuning process on the GANIT-SFT dataset, which comprises approximately 11,000 Bengali math problems with chain-of-thought reasoning, utilizing <think> and <answer> tags for structured output.
When to Use This Model
This model is ideal for applications requiring accurate and concise mathematical problem-solving in Bengali. It serves as a strong foundation for further specialized training, particularly for those interested in reinforcement learning enhancements (e.g., GRPO/CGRPO) for even better performance, as noted by the developer.