Qwen2.5-3B-Instruct-Reasoning-for-math

🧠 Qwen2.5-3B-Instruct-Reasoning-for-math

Qwen2.5-3B-Instruct-Reasoning-for-math is a fine-tuned version of Qwen2.5-3B-Instruct, specialized for mathematical reasoning in both Bangla 🇧🇩 and English 🇬🇧.

This model has been fine-tuned using Group Relative Policy Optimization (GRPO) to generate structured step-by-step reasoning chains, ideal for math education, automated tutoring, and reasoning-based chat systems.

---

📌 Model Overview

---

💡 Use Cases

This model is designed for practical applications in education, research, and language technologies:

1. 🧮 Bangla & English Math Tutoring

   • Solve arithmetic problems step by step in natural language.
   • Provide clear, structured explanations to students.

2. 🤖 Educational Chatbots

   • Integrate into interactive tutoring platforms for automated problem-solving and explanation.
   • Perfect for schools, e-learning platforms, and mobile education apps.

3. 📚 Reasoning Benchmarking & RL Research

   • Evaluate structured reasoning capabilities using XML tags (<reasoning>...</reasoning>).
   • Ideal for academic research in RLHF and structured reasoning for LLMs.

4. 🧠 Math Reasoning Agents

   • Deploy in multi-agent systems where reasoning traceability is crucial.
   • Works well with RAG pipelines for mathematical problem solving.

---

🏋️‍♂️ Training Summary

• Framework: Unsloth + TRL
• Method: GRPO RL Fine-tuning + LoRA
• Steps: 250 RL steps
• Rewards:
  • ✅ XML format compliance
  • 🧮 Mathematical correctness
  • 📝 Structured reasoning quality

---

📈 Performance Goals

• 🟢 Improved structured reasoning using <reasoning> and <answer> tags
• 🟡 Consistent step-by-step explanations in Bangla & English
• 🟢 Efficient deployment for educational tools and reasoning agents

---

🚀 Quick Start (Transformers)

You can use this model with 🤗 Transformers in just a few lines:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "shalhamucha/Qwen2.5-3B-Instruct-Reasoning-for-math"

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto",
)

# Bangla example
prompt = tokenizer.apply_chat_template([
    {"role": "system", "content": "তুমি একটি সহায়ক গাণিতিক টিউটর। ধাপে ধাপে ব্যাখ্যা করো।"},
    {"role": "user", "content": "মারির কাছে ৩টা আপেল আছে। সে আরো ২টা কিনলো। মোট কত আপেল?"},
], tokenize=False, add_generation_prompt=True)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.95,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citation

If you use this model in your research or application, please cite:

@misc{QwenMathReasoning2025,
  author       = {Shalha Mucha Bhuyan},
  title        = {Qwen2.5-3B-Instruct-Reasoning-for-math},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/shalhamucha/Qwen2.5-3B-Instruct-Reasoning-for-math}}
}