基于 Tensorflow，仿 Scikit-Learn 设计的深度学习自然语言处理框架。支持 40 余种模型类，涵盖语言模型、文本分类、NER、MRC、知识蒸馏等各个领域

Fine-Tuning Google's Vision Transformer LoRA technique. Two different LoRA adapters are tuned for separate classification (food and human actions). A simple Gradio interface is implemented to run the inference.

A full-stack platform for designing reinforcement learning environments, running GPU-backed training, and monitoring agents in real-time.

This repository features an image sharpening pipeline using Knowledge Distillation. A high-capacity Restormer acts as the teacher model, while a lightweight Mini-UNet is trained as the student to mimic its performance.

This repo documents my participation in the Kaggle red-teaming competition focused on probing OpenAI's newly released gpt-oss-20b model for previously undiscovered vulnerabilities and harmful behaviors. The goal is to identify, document, and report up to five distinct issues, contributing to the safety and alignment of open-source AI models.

AI-Farm is a distributed deep learning training framework that enables efficient model training across multiple machines. It provides a scalable infrastructure with real-time monitoring through a web admin panel, adaptive task distribution, and support for both CPU and GPU training.

PyTorch implementation of DDPM (Denoising Diffusion Probabilistic Models) for image generation. Includes U-Net with attention, DDIM sampling, EMA training, and CelebA dataset support.

1D CNN for customer complaint text classification using PyTorch.

Developed an end-to-end LLM pipeline that extracts Python code from GitHub, builds a high-quality dataset, fine-tunes CodeGen, and performs advanced code generation with DeepSeek. Demonstrates strong capabilities in LLM training, data engineering, and model optimization.

Medical X-ray image classifier fine-tuning ResNet-18 with PyTorch.

EGM is a deep learning tool that learns your image editing style from raw/edited pairs and applies it to new images. It uses a Pix2Pix (conditional GAN) architecture.

Fine-Tuning Google's Vision Transformer LoRA technique. Two different LoRA adapters are tuned for Separate Classification (Food and Human actions). A simple Gradio interface is implemented to run the Inference.

Whatsapp Assistant using FastAPI, Google Gemini, WhisperAI, Twilio and Ngrok

Data processing, Machine Learning codes and Training scripts on Jean Zay High Performance Computing (HPC) remote GPUs

Kaggle Competition for MAP Charting Student Math Mistunderstanding