| # Saved Experiments |
|
|
| Use this folder for experiment snapshots you want to keep in git. |
|
|
| Everything in the repo's normal output locations stays ignored by default: |
|
|
| - `results/` |
| - `data/` |
| - `activations/` |
| - `vertex_downloads/` |
| - generated `vertex_jobs/runs/*.yaml` |
|
|
| If an experiment is worth preserving, copy or move its files into a dedicated |
| subfolder here, for example: |
|
|
| ```text |
| saved_experiments/ |
| iman-2026-04-03-last-token-rerun/ |
| README.md |
| run_config.yaml |
| results/ |
| data/ |
| activations/ |
| ``` |
|
|
| Suggested contents for each saved experiment folder: |
|
|
| - `README.md`: short note on what the run was and why it matters |
| - `run_config.yaml`: the exact Vertex config or local config used |
| - `results/`: selected result artifacts |
| - `data/`: the dataset snapshot for that run, if needed |
| - `activations/`: only if you intentionally want them in git |
|
|
| Recommended workflow: |
|
|
| 1. Run experiments normally using the ignored output folders. |
| 2. Decide which runs are worth keeping. |
| 3. Copy the exact files you want to preserve into `saved_experiments/<owner>-<name>/`. |
| 4. Commit only that saved snapshot. |
|
|
| Recommended naming: |
|
|
| - Vertex run configs and display names: `<owner>-<run-name>` |
| - Saved experiment folders: `<owner>-<date>-<short-description>` |
|
|
| This keeps day-to-day experiments out of git while making important runs easy |
| to preserve reproducibly. |
|
|
