Experiments¶
Experiments are the second leg of each tutorial. Keep them small enough to repeat on a limited budget while preserving interfaces that later methods can replace.
Experiment tiers¶
| Tier | Purpose | Run mode |
|---|---|---|
smoke | Check that code and configuration are intact | Run automatically on every pull request |
repro | Reproduce a key curve or conclusion | Trigger manually and save a result summary |
frontier | Expensive training or large-model evaluation | Run manually or overnight with a budget record |
Directory convention¶
text experiments/<id>-<slug>/ README.md config.yaml results.md
Model weights, datasets, and complete logs do not enter Git. The repository keeps configurations, code, metrics, an index of failure examples, and environment information.