krishanwalia30/Qwen3-16bit-OpenMathReasoning-Finetuned-Merged

🚀 Harness the Power of Qwen-3 with Enhanced Reasoning and Chat! 🚀

This model is a carefully fine-tuned version of the incredible Qwen-3-8B using cutting-edge techniques with Unsloth and Parameter-Efficient Fine-Tuning (PEFT) via LoRA. It's designed to bring you the best of both worlds: the strong general capabilities of Qwen-3 with a significant boost in logical reasoning and engaging conversational skills.

We've taken the already powerful Qwen-3 and further sculpted it using a blend of the unsloth/OpenMathReasoning-mini (Chain-of-Thought split) for advanced problem-solving and the mlabonne/FineTome-100k dataset to ensure natural and fluent interactions.

🔥 Key Features:

• Enhanced Reasoning: Excels at tasks requiring logical deduction and step-by-step thinking, thanks to fine-tuning on a dedicated reasoning dataset.
• Improved Chat: Maintains and enhances the general conversational abilities of Qwen-3, making it great for interactive applications.
• Efficient Fine-Tuning: Built using the incredibly efficient Unsloth library, resulting in faster training with less memory usage.
• PEFT (LoRA) Inside: Leverages Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning, making it easier to adapt to specific tasks without full model retraining.
• Ready to Use: Seamlessly integrates with the transformers library.

🛠️ How to Get Started:

Install the necessary libraries:

pip install transformers accelerate torch

Load and use the model:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "krishanwalia30/Qwen3-16bit-OpenMathReasoning-Finetuned-Merged"

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

messages = [
    {"role": "user", "content": "Explain the Pythagorean theorem in simple terms."},
    {"role": "assistant", "content": "Okay, here's a simple explanation:"},
    {"role": "user", "content": "Now, solve for the hypotenuse if a=3 and b=4."},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = tokenizer(text, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7, top_p=0.8, top_k=20, do_sample=True)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

⚙️ Fine-tuning Details:

• Base Model: Qwen-3-8B
• Fine-tuning Framework: Unsloth
• PEFT Strategy: LoRA
• Training Datasets:
  • unsloth/OpenMathReasoning-mini (COT split)
  • mlabonne/FineTome-100k
• Training Ratio: Approximately 30% reasoning data and 70% general chat data to balance capabilities.
• Training Infrastructure: Google Colab with a T4 GPU.
• Quantization during Training: Likely 4-bit quantization was employed during the fine-tuning process using Unsloth for memory efficiency. The final merged model is saved in 16-bit for broader compatibility.
• Key Hyperparameters:
  • per_device_train_batch_size: 2
  • gradient_accumulation_steps: 4
  • learning_rate: 2e-4
  • max_steps: 30
  • Optimizer: adamw_8bit
  • Learning Rate Scheduler: linear
  • Warmup Steps: 5
  • Weight Decay: 0.01
  • Seed: 3407

📊 Evaluation:

While rigorous quantitative evaluations are ongoing, initial assessments indicate a significant improvement in the model's ability to handle reasoning-based questions while maintaining strong general conversational skills. Further benchmarks and community feedback are welcome!

👨‍💻 Author:

[https://huggingface.co/krishanwalia30]

🔗 Learn More:

For a deeper dive into the fine-tuning process and the rationale behind the choices, check out the article: [https://medium.com/@krishanw30/b1a8f684c3f1].

🙏 Acknowledgements:

A big thank you to the brilliant teams at Qwen, Unsloth AI, and the creators of the OpenMathReasoning-mini and FineTome-100k datasets for making this project possible!

Uploaded model

• Developed by: krishanwalia30
• License: apache-2.0
• Finetuned from model : unsloth/Qwen3-8B-unsloth-bnb-4bit

This qwen3 model was trained 2x faster with Unsloth and Huggingface's TRL library.