| name | run-paper |
| description | Run a paper's training or evaluation in the gdl-papers repo. Use when the user says "train <paper>", "evaluate <paper>", "run <paper> with lr/epochs/...", or asks to sweep a hyperparameter. Reads the paper's README ## Run + ## Hyperparameters, builds the command, forwards any overrides to CLI args, runs it, and reports the result. NOT for explaining repo layout (that's gdl-repo-map).
|
run-paper
Generic runner for any paper under papers/NN-name/. The paper owns its commands
and knobs (in its README + train.py); this skill just reads them and runs.
Steps
- Resolve the paper folder:
papers/<paper>/ (e.g. 00-template).
- Read
papers/<paper>/README.md — the ## Run section gives the literal
train/evaluate commands; ## Hyperparameters lists the tunable CLI args.
- Train ("train
<paper>" [+ overrides]):
- Start from the README's train command.
- Map colloquial overrides to CLI args: "lr 1e-4, 50 epochs" →
--lr 1e-4 --epochs 50.
Only use args listed in ## Hyperparameters.
- Run via
uv run python papers/<paper>/train.py [args].
- Report the final loss and the printed
run_id.
- Evaluate ("evaluate
<paper> run <id>"):
- Run
uv run python papers/<paper>/evaluate.py --run <id>.
- Report the metric.
Rules
- Always use
uv run (never bare python).
- Never invent hyperparameters not in the paper's
## Hyperparameters.
- Don't hardcode any paper's schema here — read it from the README each time.
