Imported from chelmooz/RAG-Harvard-IT-teacher (
experimental/AGENTS.md). Install upstream withnpx skills add chelmooz/RAG-Harvard-IT-teacher --skill experimental. Copyright stays with the author.
Experimental – Future Work and Fine‑Tuning Playground
Purpose: Holds prototype implementations, exploratory notebooks, and emerging features that are not yet part of the stable stack.
Structure
experimental/
├── README.md # Index of ongoing experiments
├── fine_tuning/ # LoRA / SFT training scripts and configs
│ ├── config.yaml
│ └── train.py
└── drafts/ # Unpublished design docs and architecture sketches
└── design‑sketch.md
Where to Look
| Task | Location | Notes |
|---|---|---|
| Propose a new experiment | experimental/README.md |
Add a bullet with short title + link to folder |
| Run a fine‑tuning job | experimental/fine_tuning/train.py |
Follow config.yaml hyper‑parameters |
| Review upcoming design | experimental/drafts/ |
Designs are mutable; flag WIP in title |
Code Map
- Experiment registry:
experimental/README.md→ bullet list + links. - Training script:
experimental/fine_tuning/train.py→ loadsconfig.yaml. - Design sketches:
experimental/drafts/*.md→ markdown proposals.
Conventions
- All experiments must include
status: draftfront‑matter tag. - Naming convention:
<topic>-<owner>-<date>(e.g.,rag‑summarization‑john‑2026-09-01). - Keep experiments self‑contained; avoid importing from
backend/orfrontend/unless gated behind a feature flag. - Delete or graduate experiments after 90 days of inactivity.
Anti‑Patterns
- Persisting experimental data in
backend/orfrontend/without a deprecation plan. - Using production‑grade scripts (e.g., LoRA trainer) without proper testing harness.
- Mixing stable code with raw prototype code in the same module.
Commands
# List registered experiments
grep -E '^\s*-\s' experimental/README.md
# Run a specific fine‑tuning job (example: summarization)
cd experimental/fine_tuning
python train.py --config config.yaml --name summarization
Notes
- Experiments are not part of the release pipeline; they are evaluated manually.
- When an experiment reaches
status: stable, move its folder to the appropriate top‑level directory and add anAGENTS.mdthere. - Share experimental designs on the team channel before committing large data files.