Projects

Due to the tiny bug in the interface, please click IMAGE or TITLE to view the project details. I will fix this bug as soon as possible. Sorry for the inconvenience.

ChatLLM: From-Scratch LLM Training and Post-Training Stack

Built an end-to-end LLM stack from scratch, covering scaling law experiments, compute-optimal model sizing, pre-training, supervised fine-tuning, and GRPO-based post-training. Trained a ChatGPT-style conversational model and deployed an interactive demo to showcase instruction following, reasoning, and multi-turn generation.

LLM Pre-Training Scaling Laws SFT GRPO Mixed Precision Training ZeRO-2

Yuyang Zhang

Level: Advance

⏱ 50 hours

Qwen3-VL Inference

Built a Qwen3-VL model from scratch in PyTorch, loaded the 14B (224x224) model, fine-tuned it with LoRA for specific tasks, and developed a Gradio app to showcase its capabilities.

Qwen3-VL PyTorch LoRA DeepStack M-RoPE

Yuyang Zhang

Level: Intermediate

⏱ 15 hours

Qwen3-VL Inference

Large Language Model Multi-Modality Fine-Tuning

Built a Qwen3-VL model from scratch in PyTorch, loaded the 14B (224x224) model, fine-tuned it with LoRA for specific tasks, and developed a Gradio app to showcase its capabilities.

Qwen3-VL PyTorch LoRA DeepStack M-RoPE

Yuyang Zhang

Level: Intermediate

⏱ 15 hours

PaliGemma Inference and LoRA Fine Tuning

Large Language Model Multi-Modality Fine-Tuning

This project involves modeling a multi-modal large language model called PaliGemma, which integrates vision and language tasks. First we perform inference using the pre-trained PaliGemma model. Then, we fine-tune the model using LoRA (Low-Rank Adaptation) for specific applications, for example Receipt OCR Extraction. By the end, will implement a Gradio web interface to demonstrate the fine-tuned model's capabilities.

LoRA PyTorch Gradio Grouped Query Attention RoPE KV Cache

Yuyang Zhang

Level: Entry

⏱ 8 hours