GL3MON/Qwen3.5-2B-Tamil-TAPO

Qwen3.5-2B Tamil TAPO: Specialized Tamil Math Problem Solver

This model, GL3MON/Qwen3.5-2B-Tamil-TAPO, is a 2.3 billion parameter variant of the Qwen3.5-2B architecture, specifically adapted for solving mathematical word problems presented in Tamil. It incorporates a merged LoRA (Low-Rank Adaptation) that was trained using a combination of Supervised Fine-Tuning (SFT) and a local adaptation of TAPO (Translation-Augmented Policy Optimization).

Key Capabilities

• Tamil to English Translation for Math Problems: Automatically translates Tamil mathematical questions into English within <english_translation>...</english_translation> tags.
• English Reasoning: Performs the core mathematical reasoning in English, ensuring consistency and leveraging the base model's capabilities.
• Structured Answers: Provides the final answer in a \boxed{...} format after the English reasoning.
• Specialized Training: Fine-tuned on 2,000 Tamil GSM8K training questions for SFT and 7,345 translated GSM8K questions for RL, focusing on mathematical problem-solving.

Good for

• Solving Tamil Mathematical Word Problems: Ideal for applications requiring the interpretation and solution of arithmetic and reasoning problems presented in the Tamil language.
• Multilingual Educational Tools: Can be integrated into educational platforms or tools that assist Tamil-speaking students with math.
• Research in Multilingual LLMs: Useful for researchers exploring cross-lingual transfer and specialized task adaptation in large language models, particularly for low-resource languages or specific domains like mathematics.