Welcome to my GitHub profile! I'm a second-year student at IIT Indore with a deep passion for Machine Learning, particularly in Generative AI. My work focuses on creating and optimizing AI systems that leverage generative models to tackle complex, real-world challenges innovatively and efficiently. I’m also an enthusiastic CP/DSA solver, constantly improving my problem-solving skills.
- Current Education: B.Tech. at IIT Indore
- Interests: Machine Learning, Deep Learning, Generative AI, Natural Language Processing, Computer Vision, and Backend Development.
- Current Focus: Enhancing my expertise in AI by building scalable ML systems and contributing to impactful projects.
- Objective: Developed a two-step method for LLM (Large Language Model) distillation optimized for low VRAM environments.
- Highlights:
- Stage 1: Saved logits generated by the teacher model to reduce runtime memory usage.
- Stage 2: Performed distillation using KL-divergence between the student model and pre-saved logits.
- Enabled efficient model training for users with limited hardware resources without sacrificing performance.
- Built using PyTorch, Hugging Face Transformers, and Accelerate.
- Objective: This [roject was created to make work of Competitive Programmers easy.
- Highlights:
- Used DeepSeekv2 as the main LLM for solving problems.
- Incorporated phi-3 for prompt engineering to make work of programmers fasters.
- Also tried out various techniques like ORPO for finetuning the LLM.
- Programming: Python, TensorFlow, PyTorch, C++, Django
- ML Frameworks: TensorFlow, PyTorch, Hugging Face Transformers, Diffusers
- Data Handling: Pandas, NumPy, OpenCV
- Deployment: Django, Flask, Accelerate (for distributed ML training)
- Reinforcement Learning and advanced Text-to-Image models such as Stable Diffusion.
- Optimizing model deployment for resource-constrained environments.
- LinkedIn: Harshvardhan Choudhary
- GitHub: You're already here! 🎉 Feel free to explore my repositories.
Thank you for visiting! If you have ideas for collaboration or feedback, feel free to reach out.