skilltraining-project / skill-training

Skill Training 
 Let your agent become a real expert in any domain. 
 🌐 Project page ·
 📄 PDF ·
 📑 arXiv ·
 📰 News ·
 🚀 Quick start ·
 📊 Results ·
 📖 Guide ·
 📝 Citation 
 English · 简体中文 
**Skill Training teaches an agent professional craft from finished human work, with the model weights frozen.** It corrupts a professional screenplay into an outline, asks the agent to rebuild it, and turns the gap between the draft and the original into readable skills.
- **No labels, no rewards.** The human artifact is its own teacher. Training needs no new annotations, preference data, or external reward model.
- **Skills you can read and edit.** What the agent learns is a library of small Markdown cards, and every update is a Git commit you can inspect, revert, or reuse.
- **Measurable gains that transfer.** Trained on 20 short-drama novels, the library improves all six measured writing behaviors on 385 episodes from unseen novels, wins 49 of 75 blind reader judgments, and also helps other models.
### A different route to domain expertise
**Fine-tuning changes the model. Hand-written skills need experts. Skill Training lets the agent write its own skills by rebuilding finished human work.**
| Approach | Works with closed models | No GPU training | Readable & editable | Survives a model upgrade | No experts writing skills | No success signal needed | Learns from its own attempts |
|:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
| Fine-tuning (SFT, LoRA) | ✗ | ✗ | ✗ | ✗ | ✓ | ✓ | ✗ |
| Hand-written skills | ✓ | ✓ | ✓ | ✓ | ✗ | ✓ | ✗ |
| Skills from successful runs | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ | ✓ |
| Rules summarized from examples | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✗ |
| **Skill Training (ours)** | **✓** | **✓** | **✓** | **✓** | **✓** | **✓** | **✓** |
## News
- **2026-10** — 🎬 **In production at scale.** Skills learned with our method now run in the **top-ranked short-drama generation pipeline on TikTok** (including both screenplay generation and video generation).
- **2026-09-26** — 💻 **Code released.** The reference implementation is public, with an openly licensed example you can train end to end. See Quick start.
- **2026-07** — 📑 **Preprint on arXiv.** Read it at arXiv:2607.27557.
## Overview
Professional work contains skills that are difficult to explain as instructions. Skill Training learns those skills from existing human artifacts.
The idea is simple: compress a professional screenplay into a story outline, ask an agent to reconstruct it, and learn from the gap between its draft and the original. What improves is an external **skill library**: small Markdown files the agent reads before writing. Each update is recorded in Git, so the learned rules can be inspected, edited, and reused.
This repository is the reference implementation of **Skill Training with Corruption and Reconstruction Loop**.
| Stage | What happens | Entry point |
| :--- | :--- | :--- |
| **Corrupt** | Remove craft-level detail while preserving the story outline. | `python -m skilltrain diffus