run-experiments

name

run-experiments

description

Run IntersectionQA and IntersectionEdit experiments in this repository, including local preflight checks, dataset/config selection, experiment-suite orchestration, Vast.ai GPU instance selection, remote bootstrap, SFT/GRPO launch and monitoring, stop rules, artifact preservation, and experiment-record updates. Use when Codex is asked to prepare, launch, resume, debug, monitor, or document training/evaluation experiments for this repo.

Run Experiments

Core Rule

Run experiments through the repo's restartable scripts and manifests, not one-off manual training commands, unless you are doing a bounded diagnostic. Keep local commands behind rtk; on a Vast PyTorch image, use the image Python directly after running scripts/devops/bootstrap_vast_instance.sh.

First Files To Read

Read only what is relevant to the requested run:

• specs/research-experiment-spec.md for paper experiment scope, budgets, required reporting, and split hygiene.
• configs/overnight_experiment_suite.yaml for the current full-suite manifest.
• configs/orchestration_smoke.yaml for a cheap local orchestrator smoke.
• docs/experiments/ for the most recent dated result on the same model/task.
• references/experiment-workflow.md when the task involves Vast.ai, GPU rental, launch commands, monitoring, stop rules, or artifact preservation.

Workflow

1. Identify the experiment family: dataset report, baseline, zero-shot, SFT, reasoning-SFT, GRPO/GSPO, evaluation, or analysis.
2. Prefer an existing manifest entry or config. Add or edit configs before writing new ad hoc shell commands.
3. Run cheap local checks before renting GPU: rtk uv run python -m compileall -q intersectionqa scripts, focused tests for touched code, and an orchestrator --dry-run or tiny smoke.
4. Use public train only for optimizer updates. Derive SFT/RL inner splits with existing group-safe helpers; never train on validation or test_*.
5. For Vast runs, filter offers first, sort by total hourly price, verify the live GPU/VRAM/price after creation, then bootstrap with the repo script.
6. Launch long jobs in tmux or nohup, with explicit run directories, metrics JSONL, quality eval cadence, checkpoint save cadence, and stop rules.
7. Monitor early optimizer steps, quality samples, invalid-output rate, disk, GPU memory/utilization, and budget. If GPU compute or memory utilization is below 75% during steady training, treat the run as underutilized; ideally aim for 90%+. Stop unhealthy, overpriced, or materially underutilized runs and tune the config before resuming from a checkpoint.
8. Preserve artifacts before teardown: logs, metrics, predictions, adapters, checkpoints, best-checkpoint selection, command, environment, git state, checksums, and upload paths.
9. Persist every meaningful experiment or coherent experiment set in docs/experiments/, even when the result is a failed canary or a negative result. Performance and utilization tuning are only one subsection of the record.
10. Update the relevant dated file with purpose, hypothesis, dataset/splits, config/manifest, exact commands, hardware, outcomes, failures, artifact locations, and follow-up decisions.

Command Patterns

List or dry-run the orchestrator:

rtk uv run python -m scripts.experiments.run_experiment_suite \
  configs/overnight_experiment_suite.yaml --list

rtk uv run python -m scripts.experiments.run_experiment_suite \
  configs/overnight_experiment_suite.yaml --run grpo_canary --with-dependencies --dry-run

Run a selected suite locally or remotely:

rtk uv run python -m scripts.experiments.run_experiment_suite \
  configs/orchestration_smoke.yaml --with-dependencies

On a bootstrapped Vast instance, omit rtk and use image Python:

cd /root/IntersectionQA
python -m scripts.experiments.run_experiment_suite \
  configs/overnight_experiment_suite.yaml --run grpo_canary --with-dependencies

Escalation Rules

Pause before spending GPU budget when local validation fails, split leakage is unknown, dataset artifacts are stale, Vast pricing is outside budget, expected artifacts are not configured, or stop conditions are ambiguous.

Do not write secrets, HF tokens, private SSH keys, or live credentials into repo files. Historical instance IDs can stay in dated experiment records; new credentials should stay out of docs and logs.