mdhamidhosen/baamr-math-qwen2.5-1.5b-12k
mdhamidhosen/baamr-math-qwen2.5-1.5b-12k is a 1.5 billion parameter Qwen2.5-Instruct model fine-tuned by mdhamidhosen for bilingual (Bangla and English) secondary-school mathematics. It is specifically optimized for solving math problems step-by-step, grounded in the NCTB Classes 6–10 mathematics textbooks. This model excels at providing detailed solutions to mathematical problems in the language of the query, making it suitable for educational applications.
Loading preview...
Model Overview
This model, mdhamidhosen/baamr-math-qwen2.5-1.5b-12k, is a fine-tuned version of Qwen/Qwen2.5-1.5B-Instruct, specifically adapted for solving secondary-school mathematics problems in both Bangla and English. It was trained using a LoRA adapter on the BaAMR-Math-12K dataset, which comprises 12,000 records covering all 61 chapters of the NCTB Classes 6–10 mathematics textbooks.
Key Capabilities
- Bilingual Math Problem Solving: Answers mathematics problems in either Bangla or English, matching the language of the question.
- Step-by-Step Solutions: Provides detailed, step-by-step solutions, concluding with the final answer in the format 'Answer: ...' or 'উত্তর: ...'.
- Educational Focus: Grounded in the curriculum of Bangladeshi secondary schools (NCTB Classes 6–10).
Performance and Limitations
Evaluated on a held-out test set, the model achieved an overall accuracy of 49.60%, outperforming the zero-shot base model (48.60%). It showed varying accuracy across question types and classes, with short_answer questions performing best at 67.92%. Due to its 1.5B parameter size and training on template-generated problems, it is not intended as a general math model for complex problems like those found in GSM8K or olympiads. Users should verify answers, especially for longer multi-step items, and note the observed gap in performance between Bangla and English questions.