Training/Fine-Tunning Foundation model

[tirato]

When I was tring ti run the dbdemos for llm-rag-fine-tuning I'm stuck on one of the first lines of code. When trying to execute the function: databricks.model.foundation_model.create(...) with the Mistral_7B_instruct_v2.0 it is failing with the following error:

>DatabricksModelTrainingRequestError: Context creation failed: User is not authorized to access this feature

>MAPIException: Error 500: Context creation failed: User is not authorized to access this feature

Any idea why it is popping out like that? Can it be due to being on a PublicPreview feature and needing to request access for it?

[Kaniz_Fatma]

Hi @tirato,

• Ensure that the service principal you’re using has the necessary permissions. You might need to add the service principal to the list of authorized users in the Account Console.
• After creating the access connector, an API call is made to the Databricks account. The service principal making that call needs to be known in the Databricks account¹.
• If you’re accessing Databricks endpoints with just the access token (e.g., for CI/CD workflows), add...².

1. Role Assignments:

   • In Azure Portal, navigate to Azure Databricks > Azure Databricks Service > Access control (IAM).
   • Add a role assignment for your service principal, granting the necessary roles.
   • Use the service principal to obtain the token for authentication².

2. Check Workspace Configuration:

   • Verify that your Databricks workspace configuration is correct. Ensure that the necessary parameters (such as azure_workspace_resource_id, host, and auth_type) are ...³.

If you’re still facing issues, consider reaching out for further assistance. 😊

[tirato]

Currently using my account and the command dbutils.notebook.entry_point.getDbutils().notebook().getContext().apiUrl().get().apiToken().get() to get my current databricks token to access this function.  And my user just doesn't have admin permissions over the workspace. 
It doesn't seem to be a role/permission issue. I'm able to access other PublicPreview features (like LLM RAG with vector search), just this foundation model training in particular is causing trouble.