|
Software Engineering undergraduate · Fudan University
From words to weights, from code to core, |
| Focus | What I'm working through | Resources |
|---|---|---|
learning |
Tracing a Transformer, from tokens to logits. | Karpathy's series · GPT / nanoGPT · BPE |
practicing |
Fine-tuning models and inspecting their mistakes. | PyTorch · Hugging Face · classification experiments |
foundations |
How programs represent, move, and compute with data. | CS61 series · CS170 · CSAPP |
next |
Measuring and optimizing training and inference. | CS336 · LLM systems · AI infra |
Reading & building notes
- Programming and systems foundations: CS61A / CS61B / CS61C, CS170, and CSAPP. Public coursework: CS61A and CS61B.
- Models from scratch: learning through Karpathy's micrograd, makemore, GPT / nanoGPT, tokenizer / BPE, and GPT-2 reproduction series.
- Training and post-training: experimenting with PyTorch and Hugging Face while learning more about LLM fine-tuning and post-training.
- Further down the stack: building foundations for CS336 and gradually exploring inference, memory, runtimes, and distributed training.
This map covers ongoing study and future plans.
|
01 / TRAIN THE MODEL
From LeNet to DeepConvNet: image classification, regularization, and hyperparameter experiments. PyTorch image classification
|
02 / RUN THE EXPERIMENT
Emotion classification with feature extraction, fine-tuning, and a closer look at the errors. Hugging Face fine-tuning
|
|
03 / BUILD THE INTERFACE
Four games in one Java platform, with a JavaFX GUI and a Lanterna terminal interface. Java JavaFX & Lanterna
|
04 / MAKE IT DO SOMETHING
Read the tiles, track the game, choose a move. Screen recognition connected to a strategy engine. Python OpenCV
|
More experiments, tools, and coursework
| Repository | What I built |
|---|---|
| Tiny Transformer | A Transformer forward pass with a 32-token vocabulary and explicit tensor shapes. |
| Fourier Explained | An interactive Fourier transform explainer with adjustable parameters. |
| VR Classroom | A virtual classroom demo built with Three.js / WebXR. |
| Sorting Experiment | Merge sort, insertion sort, and benchmarks for small subarrays. |
| Mathematical Modeling Skill | A reusable workflow for mathematical modeling, experiments, and result verification. |
Updated daily. Language shares reflect code volume in public repositories, excluding this profile; they do not measure proficiency. Streaks use the past-year calendar, and the current streak may end yesterday while today is still empty.
Feel free to open an issue in a project to discuss the implementation, experiment results, or anything that could use a clearer explanation.