How do I suggest adding a model to the Azure models catalog?

Models
6

I thought you were somewhere along Path B as described below…

Use two paths in parallel:

  1. Submit inside Azure AI Foundry via the Request a model button.
  2. File the same request with Hugging Face using their “Request a model addition” guidance or the HF forum category for the Azure model catalog.

Both are expected. Azure surfaces a Hugging Face collection that HF curates; Microsoft exposes it in the catalog. (Microsoft Learn)

Background you need

  • Who owns what: Azure AI Foundry/AML shows a Hugging Face collection. Microsoft’s docs state “Hugging Face creates and maintains models listed in this collection.” So model-onboarding into that collection goes through HF, while Azure provides the in-portal request route. (Microsoft Learn)
  • How it appears in Azure: Models listed as “Hugging Face” come from an Azure ML registry named HuggingFace. Azure’s guide makes this explicit. (Microsoft Learn)

Path A — Request from inside Azure AI Foundry

  1. Go to the Model Catalog (ai.azure.com → Catalog).
  2. Search your model name. If it is not present, you will see Request a model. Click it.
  3. Fill the form with your Hub URL and details. Microsoft’s page documents the button and the form. (Microsoft Learn)

Path B — Ask Hugging Face to add it to the Azure collection

Use any or all of these channels. They route to the same HF curation pipeline.

  • HF docs: “Request a model addition in the Hugging Face collection on Azure.” It explains the process and the acceptance checks. (Hugging Face)
  • HF forum category: Azure ML Studio Model Catalog for requests and escalations. (Hugging Face Forums)
  • HF GitHub repo landing: points to the HF-on-Azure docs set if you prefer to reference a canonical HF property. (GitHub)

What HF typically checks before listing

Prepare your repo so it clears the common gates:

  • Framework: Model should load with standard libraries HF supports on Azure (e.g., Transformers, Diffusers, Sentence-Transformers). (Hugging Face)
  • Task mapping: Tag your Hub repo with a supported task Azure recognizes in the catalog (e.g., chat-completion/text-generation, embeddings). HF’s Azure docs describe supported tasks and how the “Request to add” UI appears when a model is not yet listed. (Hugging Face)
  • Weights: Prefer .safetensors artifacts. This is the default safe format across HF inference stacks and aligns with cloud security guidance. (Hugging Face)
  • No trust_remote_code: Models must load with standard classes. Requiring trust_remote_code=True is discouraged for managed cloud listings due to code-execution risk. HF docs emphasize the risk of trust_remote_code. (Hugging Face)

Practical files to include in the repo card and artifacts:

  • config.json, tokenizer.json, tokenizer_config.json, generation_config.json, and .safetensors weights.
  • Clear model card: license, intended use, datasets, evals, prompt/chat template.
    These reduce catalog ingestion and deployment failures. Azure’s registry listing uses the Hugging Face model ID you supply. (Microsoft Learn)

What to submit (use everywhere: Azure form, HF docs form, HF forum)

Copy this, replace metrics as needed.

  • Model ID: jpacifico/Chocolatine-2-14B-Instruct-v2.0.3 (Hub link) (Hugging Face)

  • Task: chat-completion / text-generation

  • License: Apache-2.0 (or the exact license in your card)

  • Artifacts: .safetensors weights; config.json; tokenizer*.json; generation_config.json

  • Security: loads with vanilla Transformers; no trust_remote_code

  • Why add it: French-specialized LLM with strong performance and visible demand

  • Evidence:

    • HF Space Leaderboard LLM FR for independent French benchmarks. (Hugging Face)
    • COLE paper (Laval University, Oct 2025) evaluating French NLU; reference in your note where Chocolatine scores appear in their results. (arXiv)
    • Series traction: Chocolatine collection and downloads on HF. (Hugging Face)

  • Contact: your email/HF handle

  • Runtime notes: expected GPU class, max context, prompt template

End-to-end flow at a glance

  1. Submit Request a model in Azure. (Microsoft Learn)
  2. Post the same payload via HF “Request a model addition” and optionally the HF forum category. (Hugging Face)
  3. HF validates the repo and adds it to the HuggingFace registry that Azure surfaces. Azure users then discover and deploy it from the catalog. (Microsoft Learn)

Quick QA and pitfalls

  • Q: My model isn’t in the list. What now?
    A: Use Request a model in Azure and the HF “Request a model addition” doc route. (Microsoft Learn)
  • Q: Can models that need trust_remote_code be listed?
    A: Avoid it. Use standard Transformers classes so cloud runtimes can load safely. (Hugging Face)
  • Q: Where do these models “live” once added?
    A: In Azure ML’s HuggingFace registry. The catalog entry references the Hub model ID. (Microsoft Learn)

Short, redundant checklist

  • Add .safetensors + all tokenizer files.
  • Tag with correct task and framework.
  • Remove any trust_remote_code requirement.
  • Prepare a strong model card with license and benchmarks.
  • Submit via Azure “Request a model” and via HF “Request a model addition”. (Microsoft Learn)

References

  • Explore Azure AI Foundry Models — explains the HF collection and the in-portal Request a model flow. Updated 2025-09-04. (Microsoft Learn)
  • Deploy models from Hugging Face hub to Azure ML — confirms models come from the HuggingFace registry. Updated 2025-07-17. (Microsoft Learn)
  • HF: Request a model addition — HF’s own procedure and criteria. (Hugging Face)
  • HF forum: Azure ML Studio Model Catalog — escalation and community channel. (Hugging Face Forums)
  • Your model: jpacifico/Chocolatine-2-14B-Instruct-v2.0.3 Hub page. (Hugging Face)
  • French benchmarks: Leaderboard LLM FR Space; COLE arXiv. (Hugging Face)
1 Like