HF Model License Drift Monitor

Unofficial Apify Actor for monitoring public Hugging Face model watchlists for license, gating, and model-card governance drift. It is designed for AI teams that need a scheduled early-warning feed before refreshing model assets, embedding model metadata into internal catalogs, or approving open-weight models for production use.

This Actor is not affiliated with Hugging Face. It does not download model weights, access private repositories, bypass gated access, or provide legal advice.

What it does

For each public model ID you provide, the Actor:

• Fetches public model metadata from the Hugging Face Hub API.
• Extracts license, gated status, repository SHA, last-modified time, library, pipeline tag, downloads, likes, and license tags.
• Optionally fetches the public model-card README and stores only a SHA-256 hash plus matched governance/license terms.
• Scores license and access risk with explicit reasons.
• Compares against optional previous rows to flag license, license-link, gating, model-card hash/term, repository SHA, risk-level, and adoption drift.
• Emits structured dataset rows and a run-level OUTPUT summary suitable for scheduled monitoring.

Why it is useful repeatedly

Open model choices are not one-and-done. Licenses, model cards, gating, repository SHAs, and adoption signals can change after a model has entered an internal workflow. Run this Actor on a schedule with the previous run's rows in previousSnapshots to detect drift before a downstream model refresh or compliance review.

Example input

{
  "modelIds": [
    "Qwen/Qwen2.5-7B-Instruct",
    "mistralai/Mistral-7B-Instruct-v0.3",
    "meta-llama/Llama-3.1-8B-Instruct",
    "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
  ],
  "includeReadmeSignals": true,
  "includeAllRows": true,
  "alertSeverityThreshold": 45,
  "previousSnapshots": []
}

Output fields

Each dataset row includes:

• requestedModelId, canonical modelId, url, author
• license, custom licenseName, licenseLink, licenseIdentitySha256, licenseRisk, licenseTags
• gated, gatedEuDisallowed, private, disabled
• sha, lastModified, pipelineTag, libraryName
• downloads, likes
• readmeSha256, readmeRiskTerms, readmeHighRiskTerms
• riskScore, riskReasons, alert, recommendedAction
• baselineStatus (FIRST_RUN, COMPARED_NO_CHANGE, CHANGED, or BASELINE_MISSING)
• changedSincePrevious, changes, downloadDelta, downloadDeltaPercent when a matching previous snapshot is supplied

The key-value store OUTPUT record summarizes counts, alerts, source stats, warnings, and errors.

Risk scoring caveats

The scoring is intentionally conservative and deterministic. It is a triage tool, not a legal opinion. Always verify the linked Hugging Face model card and license text before production, commercial, or regulated use.

Broadly:

• Common permissive licenses such as apache-2.0 and mit score low by default.
• Model-specific, copyleft, OpenRAIL-style, Llama/Gemma-style, unknown, or custom licenses are review-needed.
• Non-commercial, no-derivatives, research-only, proprietary, or restricted-use language raises high-risk alerts.
• Gated models raise review priority because repository access terms and approval workflows can matter even when metadata is public. Gating is not proof of license permission.

You can tune allowedLicenses, reviewLicenses, and highRiskLicenseTerms in the input.

Limitations

• Public Hugging Face metadata only. Private repositories and model files are not accessed.
• Model IDs are validated, redirects are limited and re-validated, DNS connections are pinned to validated public addresses, and API/model-card response bodies are byte-bounded.
• README/model-card scans are keyword based and may miss nuanced terms or produce false positives.
• A missing or ambiguous license does not mean a model cannot be used; it means a human should review it.
• The Actor does not replace legal, compliance, security, or procurement review.

Pricing

Buyer-unit price: $2.00 per 1,000 model result rows written to the default dataset, plus the $0.00005 run-start event. Because one run accepts at most 200 models, the 1,000-row unit normally spans multiple runs.

This Actor uses Apify Pay Per Event (PPE) pricing:

• apify-actor-start: $0.00005 per run start.
• apify-default-dataset-item: $0.002 per dataset row (Monitor row).
• Platform usage is included in the event price; users do not pay a separate platform-usage pass-through.
• A dataset row is exactly one emitted model-result row. Source statistics and drift details are fields in that row or the run OUTPUT; they are not separately billed rows.
• Fetch-error rows are billable when emitted, because they are actionable monitor results. Set includeAllRows=false to emit only alert/error rows.
• The Actor checks Actor.pushData() charge results, stops before additional source work after the dataset-item charge limit is reached, and reports intended versus actually written rows.

Use Apify's max total charge setting before starting a run if you want a hard spend cap.