dipta007/dagger-12B_SFT_GRPO

DAGGER-12B-SFT-GRPO: Distractor-Aware Mathematical Reasoning in Bangla

This 12-billion parameter model, developed by dipta007 and built upon Gemma-3-12B-Instruct, is optimized for solving mathematical word problems in Bengali, particularly those containing distracting information. It employs a two-stage training process: Supervised Fine-Tuning (SFT) followed by Group Relative Policy Optimization (GRPO).

Key Capabilities & Differentiators

• Distractor Robustness: Exhibits significantly lower accuracy drops (12.0-14.4 points) when faced with distractors compared to standard CoT models (up to 41 points).
• Executable Outputs: Generates computational graphs in JSON format, allowing for deterministic execution to derive numerical answers.
• Token Efficiency: Achieves performance comparable to larger reasoning models (e.g., Qwen 3-8B) with 89% fewer tokens per solution.
• Explicit Distractor Modeling: Identifies and labels irrelevant information within the generated computational graph.

Performance Highlights

While slightly lower in overall weighted average accuracy (69.4%) compared to Qwen 3-8B (71.4%) on combined MGSM and MSVAMP benchmarks, DAGGER-12B demonstrates superior robustness in distractor-augmented scenarios. Its token efficiency and explicit distractor handling are key advantages.

Best Practices for Usage

• Utilize temperature=0.7 and top_p=0.8 for optimal generation quality.
• Provide the full graph instructions within the user message, as demonstrated in the Quickstart section.
• Post-process the JSON output by parsing and executing the computational graph to obtain the final numerical answer.

Limitations

• Primarily designed for arithmetic word problems; generalization to other math domains (algebra, geometry) is not guaranteed.
• Performance is evaluated for Bangla; English capabilities are not assessed.
• Requires external JSON parsing and graph execution for result extraction.