experiment-run

What problem does it solve?

Ensures every computational experiment is launched, recorded, and validated so results are reproducible, discoverable, and auditable months later. It prevents lost metrics, hidden hyperparameters, and irreproducible outputs by enforcing config-driven runs, structured output directories, and immediate recording.

Core Features & Use Cases

• Config-first runs: All runtime parameters live in a single config file so scripts have no hardcoded magic values.
• Recording negotiation and formats: Interactive negotiation of scalar, intermediate, and artifact fields with exact scalars.csv and events.jsonl schemas.
• Robust run lifecycle: Pre-flight confirmation, code snapshot, run_manifest, Slurm/local job generation, incremental recording with flushes, and post-run validation producing run_checks.json.
• Use cases: Training jobs on GPU clusters with Slurm, local long-running data processing, repeated evaluation experiments with per-sample logging, and any workflow where checkpoints and metrics must be preserved.

Quick Start

Prepare a config, negotiate what metrics and per-item fields to record, confirm the pre-flight summary, then create the exp/ directory, snapshot the code, and generate and submit the job.