You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: articles/ai-studio/concepts/connections.md
+9-9Lines changed: 9 additions & 9 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -9,8 +9,8 @@ ms.custom:
9
9
- build-2024
10
10
- ignite-2024
11
11
ms.topic: conceptual
12
-
ms.date: 11/21/2024
13
-
ms.reviewer: sgilley
12
+
ms.date: 02/21/2025
13
+
ms.reviewer: meerakurup
14
14
ms.author: sgilley
15
15
author: sdgilley
16
16
---
@@ -31,30 +31,30 @@ As another example, you can [create a connection](../how-to/connections-add.md)
31
31
32
32
## Connections to non-Microsoft services
33
33
34
-
Azure AI Foundry supports connections to non-Microsoft services, including the following:
35
-
- The [API key connection](../how-to/connections-add.md) handles authentication to your specified target on an individual basis. This is the most common non-Microsoft connection type.
36
-
- The [custom connection](../how-to/connections-add.md) allows you to securely store and access keys while storing related properties, such as targets and versions. Custom connections are useful when you have many targets that or cases where you wouldn't need a credential to access. LangChain scenarios are a good example where you would use custom service connections. Custom connections don't manage authentication, so you'll have to manage authentication on your own.
34
+
Azure AI Foundry supports connections to non-Microsoft services, including:
35
+
- The [API key connection](../how-to/connections-add.md) handles authentication to your specified target on an individual basis. API key is the most common non-Microsoft connection type.
36
+
- The [custom connection](../how-to/connections-add.md) allows you to securely store and access keys while storing related properties, such as targets and versions. Custom connections are useful when you have many targets that or cases where you wouldn't need a credential to access. LangChain scenarios are a good example where you would use custom service connections. Custom connections don't manage authentication, so you have to manage authentication on your own.
37
37
38
38
## Connections to datastores
39
39
40
40
> [!IMPORTANT]
41
-
> Data connections cannot be shared across projects. They are created exclusively in the context of one project.
41
+
> Data connections can't be shared across projects. They're created exclusively in the context of one project.
42
42
43
43
Creating a data connection allows you to access external data without copying it to your project. Instead, the connection provides a reference to the data source.
44
44
45
45
A data connection offers these benefits:
46
46
47
47
- A common, easy-to-use API that interacts with different storage types including Microsoft OneLake, Azure Blob, and Azure Data Lake Gen2.
48
48
- Easier discovery of useful connections in team operations.
49
-
-For credential-based access (service principal/SAS/key), Azure AI Foundry connection secures credential information. This way, you won't need to place that information in your scripts.
49
+
-Credential-based access (service principal/SAS/key). Azure AI Foundry connection secures credential information so you don't need to place that information in your scripts.
50
50
51
51
When you create a connection with an existing Azure storage account, you can choose between two different authentication methods:
52
52
53
53
-**Credential-based**: Authenticate data access with a service principal, shared access signature (SAS) token, or account key. Users with *Reader* project permissions can access the credentials.
54
54
-**Identity-based**: Use your Microsoft Entra ID or managed identity to authenticate data access.
55
55
56
56
> [!TIP]
57
-
> When using an identity-based connection, Azure role-based access control (Azure RBAC) is used to determine who can access the connection. You must assign the correct Azure RBAC roles to your developers before they can use the connection. For more information, see [Scenario: Connections using Microsoft Entra ID](rbac-ai-studio.md#scenario-connections-using-microsoft-entra-id-authentication).
57
+
> When you use an identity-based connection, Azure role-based access control (Azure RBAC) determines who can access the connection. You must assign the correct Azure RBAC roles to your developers before they can use the connection. For more information, see [Scenario: Connections using Microsoft Entra ID](rbac-ai-studio.md#scenario-connections-using-microsoft-entra-id-authentication).
58
58
59
59
60
60
The following table shows the supported Azure cloud-based storage services and authentication methods:
@@ -82,7 +82,7 @@ A Uniform Resource Identifier (URI) represents a storage location on your local
82
82
83
83
## Key vaults and secrets
84
84
85
-
Connections allow you to securely store credentials, authenticate access, and consume data and information. Secrets associated with connections are securely persisted in the corresponding Azure Key Vault, adhering to robust security and compliance standards. As an administrator, you can audit both shared and project-scoped connections on a hub level (link to connection rbac).
85
+
Connections allow you to securely store credentials, authenticate access, and consume data and information. Secrets associated with connections are securely persisted in the corresponding Azure Key Vault, adhering to robust security and compliance standards. As an administrator, you can audit both shared and project-scoped connections on a hub level.
86
86
87
87
Azure connections serve as key vault proxies, and interactions with connections are direct interactions with an Azure key vault. Azure AI Foundry connections store API keys securely, as secrets, in a key vault. The key vault [Azure role-based access control (Azure RBAC)](./rbac-ai-studio.md) controls access to these connection resources. A connection references the credentials from the key vault storage location for further use. You won't need to directly deal with the credentials after they're stored in the hub's key vault. You have the option to store the credentials in the YAML file. A CLI command or SDK can override them. We recommend that you avoid credential storage in a YAML file, because a security breach could lead to a credential leak.
Fine-tuning customizes a pretrained AI model with additional training on a specific task or dataset to improve performance, add new skills, or enhance accuracy. The result is a new, optimized GenAI model based on the provided examples.
20
21
21
-
Fine-tuning refers to customizing a pre-trained generative AI model with additional training on a specific task or new dataset for enhanced performance, new skills, or improved accuracy. The result is a new, custom GenAI model that's optimized based on the provided examples.
@@ -27,19 +28,20 @@ Consider fine-tuning GenAI models to:
27
28
- Save time and resources with faster and more precise results
28
29
- Get more relevant and context-aware outcomes as models are fine-tuned for specific use cases
29
30
30
-
[Azure AI Foundry](https://ai.azure.com) offers several models across model providers enabling you to get access to the latest and greatest in the market. You can discover supported models for fine-tuning through our model catalog by using the **Fine-tuning tasks** filter and selecting the model card to learn detailed information about each model. Specific models may be subjected to regional constraints, [view this list for more details](#supported-models-for-fine-tuning).
31
+
[Azure AI Foundry](https://ai.azure.com) offers several models across model providers enabling you to get access to the latest and greatest in the market. You can discover supported models for fine-tuning through our model catalog by using the **Fine-tuning tasks** filter and selecting the model card to learn detailed information about each model. Specific models might be subjected to regional constraints. [View this list for more details](#supported-models-for-fine-tuning).
31
32
32
33
:::image type="content" source="../media/concepts/model-catalog-fine-tuning.png" alt-text="Screenshot of Azure AI Foundry model catalog and filtering by Fine-tuning tasks." lightbox="../media/concepts/model-catalog-fine-tuning.png":::
33
34
34
-
This article will walk you through use-cases for fine-tuning and how this can help you in your GenAI journey.
35
+
This article walks you through use-cases for fine-tuning and how it helps you in your GenAI journey.
35
36
36
37
## Getting started with fine-tuning
37
38
38
39
When starting out on your generative AI journey, we recommend you begin with prompt engineering and RAG to familiarize yourself with base models and its capabilities.
39
40
-[Prompt engineering](../../ai-services/openai/concepts/prompt-engineering.md) is a technique that involves designing prompts using tone and style details, example responses, and intent mapping for natural language processing models. This process improves accuracy and relevancy in responses, to optimize the performance of the model.
40
41
-[Retrieval-augmented generation (RAG)](../concepts/retrieval-augmented-generation.md) improves LLM performance by retrieving data from external sources and incorporating it into a prompt. RAG can help businesses achieve customized solutions while maintaining data relevance and optimizing costs.
41
42
42
-
As you get comfortable and begin building your solution, it's important to understand where prompt engineering falls short and that will help you realize if you should try fine-tuning.
43
+
As you get comfortable and begin building your solution, it's important to understand where prompt engineering falls short and when you should try fine-tuning.
44
+
43
45
- Is the base model failing on edge cases or exceptions?
44
46
- Is the base model not consistently providing output in the right format?
45
47
- Is it difficult to fit enough examples in the context window to steer the model?
@@ -53,26 +55,29 @@ _A customer wants to use GPT-3.5 Turbo to turn natural language questions into q
53
55
54
56
### Use cases
55
57
56
-
Base models are already pre-trained on vast amounts of data and most times you'll add instructions and examples to the prompt to get the quality responses that you're looking for - this process is called "few-shot learning". Fine-tuning allows you to train a model with many more examples that you can tailor to meet your specific use-case, thus improving on few-shot learning. This can reduce the number of tokens in the prompt leading to potential cost savings and requests with lower latency.
58
+
Base models are already pretrained on vast amounts of data. Most times you add instructions and examples to the prompt to get the quality responses that you're looking for - this process is called "few-shot learning." Fine-tuning allows you to train a model with many more examples that you can tailor to meet your specific use-case, thus improving on few-shot learning. Fine-tuning can reduce the number of tokens in the prompt leading to potential cost savings and requests with lower latency.
59
+
60
+
Turning natural language into a query language is just one use case where you can "_show not tell_" the model how to behave. Here are some other use cases:
57
61
58
-
Turning natural language into a query language is just one use case where you can _show not tell_ the model how to behave. Here are some additional use cases:
59
62
- Improve the model's handling of retrieved data
60
63
- Steer model to output content in a specific style, tone, or format
61
64
- Improve the accuracy when you look up information
62
65
- Reduce the length of your prompt
63
-
- Teach new skills (i.e. natural language to code)
66
+
- Teach new skills (that is, natural language to code)
64
67
65
-
If you identify cost as your primary motivator, proceed with caution. Fine-tuning might reduce costs for certain use cases by shortening prompts or allowing you to use a smaller model. But there may be a higher upfront cost to training, and you have to pay for hosting your own custom model.
68
+
If you identify cost as your primary motivator, proceed with caution. Fine-tuning might reduce costs for certain use cases by shortening prompts or allowing you to use a smaller model. But there might be a higher upfront cost to training, and you have to pay for hosting your own custom model.
66
69
67
70
### Steps to fine-tune a model
71
+
68
72
Here are the general steps to fine-tune a model:
69
-
1. Based on your use case, choose a model that supports your task
70
-
2. Prepare and upload training data
71
-
3. (Optional) Prepare and upload validation data
72
-
4. (Optional) Configure task parameters
73
-
5. Train your model.
74
-
6. Once completed, review metrics and evaluate model. If the results don't meet your benchmark, then go back to step 2.
75
-
7. Use your fine-tuned model
73
+
74
+
1. Choose a model that supports your task.
75
+
1. Prepare and upload training data.
76
+
1. (Optional) Prepare and upload validation data.
77
+
1. (Optional) Configure task parameters.
78
+
1. Train your model.
79
+
1. Once completed, review metrics and evaluate model. If the results don't meet your benchmark, then go back to step 2.
80
+
1. Use your fine-tuned model.
76
81
77
82
It's important to call out that fine-tuning is heavily dependent on the quality of data that you can provide. It's best practice to provide hundreds, if not thousands, of training examples to be successful and get your desired results.
78
83
@@ -87,7 +92,6 @@ For more information on fine-tuning using a managed compute (preview), see [Fine
87
92
88
93
For details about Azure OpenAI models that are available for fine-tuning, see the [Azure OpenAI Service models documentation](../../ai-services/openai/concepts/models.md#fine-tuning-models) or the [Azure OpenAI models table](#fine-tuning-azure-openai-models) later in this guide.
89
94
90
-
91
95
For the Azure OpenAI Service models that you can fine tune, supported regions for fine-tuning include North Central US, Sweden Central, and more.
0 commit comments