AI Hub Change Events

Stateful Hugging Face monitoring that emits changes, not copied catalog rows. Schedule it hourly or daily to turn public model and dataset metadata into a compact event feed for MLOps, AI governance, scouting, and automation.

Why pay for this instead of calling the Hub API?

The public API returns snapshots. This Actor keeps the baseline, compares runs, suppresses duplicates, classifies governance changes, calculates metric deltas and per-day velocity, and produces webhook-ready events with stable IDs and severity. No GPU and no LLM calls are used, so behavior and cost are deterministic.

Events

• REPOSITORY_DISCOVERED — a rolling query starts matching a new model or dataset.
• LICENSE_CHANGED — license added, removed, or changed; restrictive/unknown targets are critical.
• ACCESS_CHANGED — gated or private status changed.
• REVISION_CHANGED — repository SHA changed.
• METADATA_CHANGED — pipeline task or library changed.
• DOWNLOAD_TREND — thresholded download movement, including delta, percent, and per-day velocity.
• LIKE_TREND — thresholded like movement.

The first default run creates a silent baseline. Later runs emit only transitions. Set bootstrapMode to emit-discovered when initial discovery events are desired.

Example input

{
  "resources": [
    {"type": "model", "id": "Qwen/Qwen2.5-0.5B-Instruct"},
    {"type": "dataset", "id": "HuggingFaceH4/ultrachat_200k"}
  ],
  "discoveryQueries": [
    {"type": "model", "search": "function calling", "sort": "lastModified", "limit": 25}
  ],
  "minDownloadDelta": 1000,
  "minDownloadGrowthPercent": 10,
  "minLikeDelta": 10,
  "maxItems": 100
}

Example event

{
  "eventId": "f6df8f9b06d736049752",
  "eventType": "LICENSE_CHANGED",
  "severity": "critical",
  "source": "huggingface_hub",
  "resourceType": "model",
  "resourceId": "org/model",
  "summary": "License changed from apache-2.0 to other",
  "changes": {"license": {"from": "apache-2.0", "to": "other"}},
  "current": {"license": "other", "gated": false, "sha": "..."}
}

Scheduling and state

Use an Apify schedule with the same default key-value store. State lives under AI_HUB_CHANGE_STATE; choose a separate stateKey for each logically independent monitor. Dataset output contains events only. The OUTPUT key contains the run summary.

State is updated only after event delivery succeeds. A stable eventId hashes the transition itself, making downstream deduplication straightforward.

Limits and responsible use

• Public Hugging Face Hub data only; no model weights, private repositories, or personal-data enrichment.
• Optional token only raises Hub API limits; it does not enable private-repository monitoring.
• Download counts are Hub-provided rolling metrics, not lifetime counters. Decreases are emitted as trend changes rather than treated as errors.
• License classification is a monitoring signal, not legal advice.
• Maximum 500 unique repositories per run.

Local development

python -m venv .venv
. .venv/bin/activate
pip install -r requirements-dev.txt
pytest -q

Run locally with Apify storage:

mkdir -p storage/key_value_stores/default
printf '%s' '{"resources":[{"type":"model","id":"Qwen/Qwen2.5-0.5B-Instruct"}],"bootstrapMode":"emit-discovered"}' > storage/key_value_stores/default/INPUT.json
APIFY_LOCAL_STORAGE_DIR=./storage python -m src.main

Suggested Store pricing

Pay per emitted event: $3 / 1,000 events plus $0.005 per Actor start, platform usage included. Baseline rows and unchanged checks should not be billable. This aligns payment with the product's value: actionable changes, not API rows fetched.