Find an app
Account
Toggle menu

Use in your browser

Validate Croissant metadata

Validate a hosted or pasted Croissant document with the real mlcroissant engine: every field source resolved to a declared file, cross-RecordSet references checked, and a bounded reachability probe over the files the document declares. Compare two revisions of one dataset to see which conformance findings appeared or cleared.

Guest includes 50 lifetime app requests. Create a free account for 100 app requests each calendar month. One allowance across every app, in your browser and your agent.One allowance across every app, in your browser and your agent. Manage usage

Croissant Dataset Check

Connect agent

Guest access is available after setup. Full account linking verified by PowMCP: Claude only.

Check

Validate one hosted or pasted Croissant document with mlcroissant, then probe the files it declares for reachability.

Usually finishes in a few seconds; the reachability probe takes most of that time.

Croissant source

Only the active source is sent. You can switch without losing what you entered.

A directly hosted public Croissant JSON-LD document.

Allowed range: 1–10 · whole numbers

Left empty, up to 10 declared files are probed with HEAD or range requests.

Add Croissant Dataset Check to your agent

Connect this app on its own. Add other PowMCP apps whenever your agent needs another job done.

Open the apps workspace A glowing app tile clicking into a slot on a dark agent device, with the PowMCP app box behind it

Connect directly

Choose your agent. Each button opens a new tab with only this app's endpoint ready to add.

Use another agent

Copy one ready-to-send prompt for any agent that supports remote MCP connections.

Manual endpoint, JSON, and terminal commands

For clients that require manual configuration, use this app-only endpoint, its JSON entry, or one terminal command.

ChatGPT: enable Developer mode in Settings → Security and login, then add this MCP endpoint from the Plugins page. Availability depends on your account and workspace policy. PowMCP has not yet verified ChatGPT account linking. OpenAI setup guide

https://powmcp.com/ml-dataset-metadata-check/mcp
{
  "mcpServers": {
    "powmcp-ml-dataset-metadata-check": {
      "type": "http",
      "url": "https://powmcp.com/ml-dataset-metadata-check/mcp"
    }
  }
}

Terminal agents add this app with one command:

Claude Code
claude mcp add --transport http powmcp-ml-dataset-metadata-check https://powmcp.com/ml-dataset-metadata-check/mcp
Codex CLI
codex mcp add powmcp-ml-dataset-metadata-check --url https://powmcp.com/ml-dataset-metadata-check/mcp
Gemini CLI
gemini mcp add --transport http powmcp-ml-dataset-metadata-check https://powmcp.com/ml-dataset-metadata-check/mcp

Manage, disable, or remove this connection in your agent's own MCP settings. PowMCP does not label an external connection as installed without confirmation from that client.

Proof

What it does

Validate a hosted Croissant ML-dataset metadata doc with the real mlcroissant structure-graph engine, plus a bounded distribution-reachability probe

Validate one public Croissant (MLCommons) dataset metadata JSON-LD document against the pinned mlcroissant engine (resolving Field sources, cross-RecordSet references, and duplicate ids into a real structure graph and probing declared distributions for reachability) or compare two same-lineage Croissant docs across dataset versions, with a clear structural-metadata-versus-data-correctness boundary.

Category
Developer Tools

Request cost1 completed result

Guests get 50 lifetime app requests. Free accounts get 100 each calendar month.

Tools in this app

01

Croissant Dataset Metadata Check

croissant_check

Validates one directly hosted public Croissant (MLCommons) ML-dataset metadata JSON-LD document (a HuggingFace /api/datasets/<id>/croissant endpoint, a Kaggle/OpenML export, or a raw croissant.json) or a pasted Croissant JSON-LD document, against the pinned mlcroissant engine, which builds a real rdflib+networkx structure graph (metadata-only: it resolves every Field source to a declared distribution, resolves references/foreign-keys across RecordSets, rejects duplicate ids, and checks source-versus-value and dataType/arrayShape: it does NOT download the dataset). It then adds a bounded HEAD/range reachability probe over the declared distribution contentUrls (capped, never downloading a distribution). Returns conformance pass/fail at the detected Croissant version, the mlcroissant version, a structural inventory, per-field source resolution, cross-RecordSet reference resolution, duplicate-id findings, and per-distribution reachability, with absence encoded as distinguished states (not-a-croissant-document vs croissant-invalid; version-not-detected vs unsupported-version; distribution-declared-but-unreachable vs no-content-url; field-source-unresolved vs field-source-absent). Use for one-document conformance and diagnosis: 'is this valid Croissant', 'will it pass conformance', 'do the declared files still resolve'. A pass means only that the document is structurally valid Croissant and internally consistent and that probed distributions were reachable, never that the dataset is correct, unbiased, license-clean, or that its bytes match the fields row by row. Validate is sub-second and the bounded probe dominates, so a call typically returns in a few seconds and always well under the gate: tell the user before calling.

02

Croissant Dataset Metadata Compare

croissant_compare

Runs the identical pinned mlcroissant validation over exactly two versions or spec revisions of the same dataset's Croissant document, baseline first and revised second, sequentially under one shared deadline, then reports which conformance findings regressed or were resolved. Each source is a hosted HTTP(S) Croissant URL or pasted JSON-LD. Select this whenever the user supplies two Croissant documents and asks what conformance changed, whether a revision regressed the metadata, or which version has fewer machine-checkable findings. The tool computes comparability from the required canonical dataset URL and returns an explicit not-comparable result with no diff or ranking when either URL is missing or the normalized URLs differ. Returns per-document dataset name and canonical URL, sha256, detected Croissant version, conforms status, structural inventory, and finding counts; a comparability note; an observed spec-version delta when the two conformsTo versions differ; a difference table of findings introduced (regressions) and resolved keyed by category (field-source, RecordSet reference, duplicate id, constraint); and, only for a comparable pair, a ranking by conformance state then fewer findings. The ranking orders machine-checkable conformance only: it never means one document's dataset is truer, more complete, or higher quality, and human review remains required for both.

Croissant Dataset Check questions

Scope and boundaries for Croissant Dataset Check.

01What can I check with Croissant Dataset Check?

Validate the MLCommons GPT-3 Croissant metadata.json sample with distribution probing enabled, or compare its v1.0.1 and v1.1.0 documents as baseline and revision for the same canonical dataset URL.

02How should I read the result?

Read documentState, versionState, and checkState before conforms. Conforms true means mlcroissant validated the Croissant structure graph; distribution reachability is a separate bounded probe. Fix findings, unresolved fields or references, and duplicate IDs before unreachable distributions. Use comparison differences or ranking only when comparable is true, and treat them as incomplete when truncated is true.

03What are its limits?

Validation builds the metadata structure graph but does not download or validate dataset records; distribution checks use bounded HEAD or range probes. Results do not establish data correctness, completeness, bias, license clearance, or row-by-row agreement with declared fields, and comparisons require the same normalized canonical dataset URL.