dipta007/GanitLLM-0.6B-SFT
GanitLLM-0.6B-SFT by dipta007 is a 0.6 billion parameter causal language model, fine-tuned from Qwen3-0.6B with Supervised Fine-Tuning on the GANIT dataset. This model specializes in Bengali mathematical reasoning, demonstrating significant accuracy improvements on benchmarks like Bn-MGSM and Bn-MSVAMP. It is optimized for resource-constrained deployments and excels at generating concise, Bengali-based mathematical solutions.
Loading preview...
GanitLLM-0.6B-SFT: Bengali Mathematical Reasoning Model
GanitLLM-0.6B-SFT is a 0.6 billion parameter causal language model developed by dipta007, specifically fine-tuned for mathematical reasoning in Bengali. Built upon the Qwen3-0.6B base model, it leverages Supervised Fine-Tuning (SFT) on the GANIT dataset, which comprises over 11,000 Bengali math problems with chain-of-thought reasoning.
Key Capabilities and Performance
This model demonstrates substantial improvements in Bengali mathematical reasoning compared to its base model, making it suitable for resource-constrained environments. Notable performance highlights include:
- +20.00 accuracy on the Bn-MGSM benchmark (improving from 8.40 to 28.40).
- +39.20 accuracy on the Bn-MSVAMP benchmark (improving from 12.20 to 51.40).
- Achieves 88.60% Bengali reasoning capability, a significant increase from the base model's 12.43%.
- Generates solutions with 79.2% fewer words on average (263 words vs. 1265 words for the base model), indicating more concise reasoning.
Training Details
The model underwent a single-stage Supervised Fine-Tuning process using the GANIT-SFT dataset. This dataset is structured with <think> tags for reasoning steps and <answer> tags for final solutions, grounding the model's reasoning in Bengali.
Good for
- Bengali mathematical problem-solving and reasoning tasks.
- Applications requiring concise mathematical explanations in Bengali.
- Deployments with limited computational resources due to its small parameter count.
For enhanced performance, users are encouraged to explore the RL-enhanced versions like GanitLLM-0.6B_SFT_CGRPO or GanitLLM-0.6B_SFT_GRPO.