dipta007/dagger-4B_SFT
dipta007/dagger-4B_SFT is a 4.3 billion parameter supervised fine-tuned model based on Gemma-3-4B-Instruct, developed by dipta007. It is designed for computational graph generation, specifically for solving mathematical problems by identifying relevant and distractor information. This model serves as a lightweight baseline and an initialization point for further GRPO training, excelling in tasks requiring structured mathematical reasoning.
Loading preview...
Model Overview
dipta007/dagger-4B_SFT is a 4.3 billion parameter language model, supervised fine-tuned from the Gemma-3-4B-Instruct base model. It specializes in generating computational graphs for mathematical problems, distinguishing between essential and distractor information within the problem statement. This model was accepted at EMNLP 2026 (Findings) and is part of a research effort to improve mathematical reasoning in LLMs.
Key Capabilities
- Computational Graph Generation: Creates structured JSON graphs representing mathematical solutions, including operations, arguments, and identification of distractor nodes.
- Distractor Awareness: Explicitly marks numerical values as
distractor: trueif they are not used in the final calculation path, enhancing reasoning transparency. - Mathematical Reasoning: Designed to solve complex math problems by breaking them down into a series of computational steps.
- Lightweight Baseline: Provides a strong performance baseline for its size, particularly useful when larger models are computationally prohibitive.
Use Cases
- GRPO Initialization: Serves as an effective starting point for further policy optimization (GRPO) training to achieve higher performance.
- Ablation Studies: Ideal for research comparing the contributions of Supervised Fine-Tuning (SFT) versus GRPO in mathematical reasoning tasks.
- Lightweight Applications: Suitable for scenarios where a smaller, efficient model for structured mathematical problem-solving is required.
Performance
While dagger-4B_SFT achieves a weighted average of 44.3 on combined MGSM and MSVAMP datasets, it shows significant improvement when used as an initialization for GRPO, with dagger-4B_SFT_GRPO reaching 47.3 (+3.0). The model's training data includes numina-math-cot-bn, which has a CC BY-NC-SA 4.0 license, requiring careful consideration for commercial use.