An AI pipeline that reads exam papers and writes high-school essays

What it does WriteGPT is a five-module pipeline built to generate Chinese Gaokao argumentative essays. It photographs an exam paper, uses EAST and CRNN to read the prompt, distills the instructions with BERT, and drafts the essay with a fine-tuned GPT-2 model. The authors claim the output hits a passing grade for an average high schooler, and they envision the result flowing straight through an external device onto a physical answer card.

The interesting bit The finetuning corpus tells the real story: after seeding the model with roughly one thousand top-scoring Gaokao essays and prose collections, the team later added the Selected Works of Mao Zedong, Chen Duxiu’s Anthology, and Lu Xun’s Collected Works to the training set. It is difficult to find a blunter instrument for tuning ideological register. The project is also triple-licensed under Apache-2.0, MIT, and the Anti-996 License—a combination that is as principled as it is unusual.

Key highlights

• Five independently trained modules (EAST, CRNN, BERT, GPT-2, DNN) wired into an inference pipeline
• 1.7 billion parameters, pretrained on ~15 GB of cleaned Chinese text and fine-tuned on ~1,000 high-scoring exam essays
• BERT summarization strips out boilerplate exam instructions like “read the following materials” before generation
• Colab demo available, though only for the text-generation module
• Explicitly tagged “for communication and popular science only,” which is one way to write a disclaimer

Caveats

• Built on TensorFlow 1.15.2 and Ubuntu 18.04.2, so the stack is aging fast
• The physical answer-card output hardware is mentioned but not actually explained in the visible README
• PyPI version sits at 0.0.1a4, suggesting the project is still very much alpha

Verdict A curiosity worth inspecting if you study Chinese NLP, stylistic mimicry, or unapologetically single-purpose pipelines. Everyone else can safely skip this 2020-era time capsule.