Golden Experiments¶

A golden experiment is a release-grade run that ships with compressionKIT for v1. Each entry is fully declarative: a YAML config under configs/, a registered training recipe, and a fixed dataset + HuggingFace repo target. The lifecycle runner (compressionkit golden run <id>) reproduces it end-to-end from a clean checkout.

Two families exist:

• codec — single-stage RVQ autoencoder (encoder → RVQ → decoder).
• two_stage — paired entropy prior on top of a parent codec. Trains a small causal-transformer prior over the codec's token stream and bundles prior_int8.tflite alongside the codec artifacts in the same HuggingFace repo.

v1 Registry¶

Reproduce one experiment¶

# 1. Fetch the dataset (MESA requires NSRR_TOKEN; PTB-XL is open).
uv run compressionkit golden run ppg-rvq-4x --skip-dataset-check  # smoke

# 2. Real run, publish to HuggingFace (set HF_TOKEN first).
uv run compressionkit golden run ppg-rvq-4x --publish

Reproduce every experiment in a modality¶

uv run compressionkit golden run-all --modality ppg
uv run compressionkit golden run-all --modality ecg

See also:

• HuggingFace testing guide — load any AmbiqAI model in five minutes.
• Deployment guide — exporting the artifacts to an Ambiq-class device.
• Methods · RVQ Autoencoder — the architecture every entry uses.