dipta007/dagger-12B_SFT

VISIONPricing:Input $0.2 / Output $0.6Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kPublished:Jan 1, 2026License:gemmaArchitecture:Transformer0.0K Featherless Exclusive Cold

dipta007/dagger-12B_SFT is a 12 billion parameter supervised fine-tuned model based on Gemma-3-12B-Instruct, developed by dipta007. It specializes in generating computational graphs for Bangla mathematical reasoning, specifically designed to handle distractor-aware problems. This model serves as a strong baseline for mathematical reasoning tasks in Bengali and is intended for use as an initialization for further policy optimization methods like GRPO.

Loading preview...

DAGGER-12B-SFT: Computational Graph Generation for Bangla Math

DAGGER-12B-SFT is a 12 billion parameter model, built upon the Gemma-3-12B-Instruct architecture, specifically fine-tuned for generating computational graphs to solve mathematical reasoning problems in Bangla. This model focuses on distractor-aware reasoning, meaning it can identify and correctly process relevant information while ignoring irrelevant numerical distractors within a problem.

Key Capabilities and Features

  • Supervised Fine-Tuning (SFT): Trained on 3,000 verified computational graph examples, providing a robust foundation for complex math problems.
  • Bangla Mathematical Reasoning: Optimized for understanding and solving math problems presented in the Bengali language.
  • Computational Graph Output: Generates structured JSON outputs representing the step-by-step solution as a graph, including operations, arguments, and distractor identification.
  • Baseline Performance: Achieves a weighted average accuracy of 66.7% on datasets like MGSM and MSVAMP, serving as a strong starting point for further development.
  • Efficient Inference: Designed for practical use, processing approximately 334 tokens per problem on average.

When to Use This Model

  • As a Research Baseline: Ideal for comparing against more advanced reinforcement learning methods, such as GRPO-enhanced variants.
  • GRPO Initialization: Functions as an effective starting point for policy optimization training (e.g., for dagger-12B_SFT_GRPO).
  • Resource-Constrained Environments: Suitable for scenarios where extensive GRPO training is not feasible due to computational limitations.
  • Studying SFT vs. GRPO: Valuable for researchers investigating the impact of supervised fine-tuning versus policy optimization on graph generation tasks.