Deposit format

Publishing a model means publishing two kinds of thing: the model file, which is large and stays out of Git, and four small files describing it, which are committed.

deposits/<setting>/<quantity>/<family>/
├── README.md       # model card, and the record's preview page
├── manifest.json   # which bytes, and what is needed to load them
├── model.json      # what the artifact is, machine-readable
└── metadata.json   # PSDI discovery metadata — sent, not uploaded

The first three are uploaded alongside the artifacts. metadata.json is what PSDI is told, not a file it receives.

model.json

Copy it from your run's model/ folder — the training run writes it. It is what makes a record loadable rather than only described, and publishing refuses a deposit without one.

For an artifact fitted before that was true, write one afterwards and set "record_origin": "reconstructed".

role says what the artifact is for:

Role	Meaning
model	answers a question; load_model serves it
feature_extractor	supplies input to another model; load_model declines it by name

manifest.json

{
  "schema_version": 1,
  "community": "data-to-knowledge",
  "artifacts": [
    {"name": "model.bin", "size_bytes": 12345, "sha256": "64 lowercase hex chars"}
  ],
  "inference_requirements": {
    "artifact_format": "describe the serializer",
    "feature_contract": "versioned-feature-schema",
    "target": "prediction_target"
  }
}

Generate every artifact entry with goldilocks-ml publish checksum PATH rather than typing sizes and digests. Names must be basenames, and cannot be README.md, manifest.json or metadata.json.

inference_requirements is free-form and is for whoever loads the file later: the serializer, the library versions that matter, the feature schema, the prediction target, and any support files.

metadata.json

Copy the closest existing deposit and check these:

Field	Purpose
access	public record and files
files.default_preview	README.md, so the record page shows the card
metadata.title	specific model and task
metadata.description	short HTML shown in search results
metadata.creators	who the citation names
metadata.contributors	other project roles
metadata.rights	the licence on the artifact
metadata.resource_type	model
metadata.version	the release version
custom_fields.dsmd	[] — see below

PSDI only stores custom_fields.dsmd keys it has registered, and drops the rest silently with a 200. None of ours are registered, so send an empty list rather than a block that looks stored and is not.

Creators and contributors do different jobs: PSDI builds the citation from creators, so do not add the whole project there.

README.md

The model card should let a colleague decide whether and how to use the model without opening anything else:

1. what it predicts, and in what units or convention;
2. how to load it, and the library versions that matter;
3. how good it is, measured on held-out data;
4. where it fails, and what it has not been tested on;
5. what the training data was;
6. checksum verification, and an unsafe-deserialization warning for a pickle.

Do not claim metrics, dataset digests or guarantees that cannot be tied to the exact released artifact.

No Markdown tables. PSDI previews the card with a plain Markdown renderer that has no table extension, so a table arrives as a wall of pipes. Publishing refuses a card containing one. Lay numbers out as aligned columns inside a fenced block, which renders anywhere. A table inside a fenced block is fine.