morel
Docs/morel — Reproduction

morel — Reproduction

A single command reproduces an experiment deterministically.

Quick reproduction (synthetic)

python -m morel train completion

This runs the full pipeline on a synthetic dataset and writes a reproducible artifact bundle under runs/<timestamp>/:

runs/<timestamp>/
├── config.yaml           # the exact Config used
├── manifest.json         # dataset, version, code hash, seed
├── metrics.jsonl         # per-step metrics
├── FIDELITY.md           # rendered from the fidelity registry
├── FIDELITY.json         # machine-readable fidelity report
├── report.md             # one-paragraph summary of the run
└── checkpoints/
    └── best.pt           # best validation checkpoint

Reproducing from a saved config

python -m morel reproduce path/to/config.yaml --items 100 --users 30 --epochs 5

The Reproduce service reads the YAML, sets the seed, validates the manifest's config_hash if present, and delegates to Experiment.run for full reproducibility.

Determinism guarantees

  • morel.core.seed.seed(value) configures deterministic seeding for torch, torch CUDA, numpy, Python random, PYTHONHASHSEED, and cuDNN. The CLI calls it at startup.
  • Resuming a run requires the saved config_hash to match the new Config; mismatched hashes raise the typed Cfg configuration error.
  • Manifest sidecar binding ensures the artifact cannot be silently consumed under a different configuration.

Programmatic

from morel.app import Reproduce

result = Reproduce(
    config_path="runs/<ts>/config.yaml",
    run_dir="runs/<ts>",
).run()

Limitations

  • The default synthetic reproduction uses random features and a small synthetic bipartite graph. Real-data reproduction requires either morel data download (Amazon-Reviews-2023 mirror) or a manually curated data/processed/<dataset> directory matching the schema.
  • See LIMITATIONS.md for the published reproducibility scope.