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-foundry/concepts/model-benchmarks.md
+42-3Lines changed: 42 additions & 3 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -20,14 +20,14 @@ author: lgayhardt
20
20
21
21
Model leaderboards (preview) in Azure AI Foundry portal allow you to streamline the model selection process in the Azure AI Foundry [model catalog](../how-to/model-catalog-overview.md). The model leaderboards, backed by industry-standard benchmarks can help you to find the best model for your custom AI solution. From the model leaderboards section of the model catalog, you can [browse leaderboards](https://aka.ms/model-leaderboards) to compare available models as follows:
22
22
23
-
-[Quality, cost, and performance leaderboards](../how-to/benchmark-model-in-catalog.md#access-model-leaderboards) to quickly identify the model leaders along a single metric (quality, cost, or throughput);
23
+
-[Quality, safety, cost, and performance leaderboards](../how-to/benchmark-model-in-catalog.md#access-model-leaderboards) to quickly identify the model leaders along a single metric (quality, safety, cost, or throughput);
24
24
-[Trade-off charts](../how-to/benchmark-model-in-catalog.md#compare-models-in-the-trade-off-charts) to see how models perform on one metric versus another, such as quality versus cost;
25
25
-[Leaderboards by scenario](../how-to/benchmark-model-in-catalog.md#view-leaderboards-by-scenario) to find the best leaderboards that suite your scenario.
26
26
27
27
Whenever you find a model to your liking, you can select it and zoom into the **Detailed benchmarking results** of the model within the model catalog. If satisfied with the model, you can deploy it, try it in the playground, or evaluate it on your data. The leaderboards support benchmarking across text language models (large language models (LLMs) and small language models (SLMs)) and embedding models.
28
28
29
29
30
-
Model benchmarks assess LLMs and SLMs across the following categories: quality, performance, and cost. In addition, we assess the quality of embedding models using standard benchmarks. The leaderboards are updated regularly as better and more unsaturated benchmarks are onboarded, and as new models are added to the model catalog.
30
+
Model benchmarks assess LLMs and SLMs across the following categories: quality, safety, cost, and throughput. In addition, we assess the quality of embedding models using standard benchmarks. The leaderboards are updated regularly as better and more unsaturated benchmarks are onboarded, and as new models are added to the model catalog.
31
31
32
32
33
33
## Quality benchmarks of language models
@@ -40,7 +40,7 @@ Azure AI assesses the quality of LLMs and SLMs using accuracy scores from standa
40
40
41
41
Quality index is provided on a scale of zero to one. Higher values of quality index are better. The datasets included in quality index are:
42
42
43
-
| Dataset Name | Leaderboard Category|
43
+
| Dataset Name | Leaderboard Scenario|
44
44
|--------------------|----------------------|
45
45
| arena_hard | QA |
46
46
| bigbench_hard | Reasoning |
@@ -62,6 +62,45 @@ See more details in accuracy scores:
62
62
Accuracy scores are provided on a scale of zero to one. Higher values are better.
63
63
64
64
65
+
## Safety benchmarks of language models
66
+
67
+
To guide the selection of safety benchmarks for evaluation, we apply a structured filtering and validation process designed to ensure both relevance and rigor. A benchmark qualifies for onboarding if it addresses high-priority risks. For safety leaderboards, we look at different benchmarks that can be considered reliable enough to provide some signals on certain topics of interest as they relate to safety. We select [HarmBench](https://github.com/centerforaisafety/HarmBench) to proxy model safety, and organize scenario leaderboards as follows:
| HarmBench (standard) | Standard harmful behaviors | Attack Success Rate | Lower values means better robustness against attacks designed to illicit standard harmful content |
72
+
| HarmBench (contextual) | Contextually harmful behaviors | Attack Success Rate | Lower values means better robustness against attacks designed to illicit contextually harmful content |
73
+
| HarmBench (copyright violations) | Copyright violations | Attack Success Rate | Lower values means better robustness against attacks designed to illicit copyright violations|
74
+
| WMDP | Knowledge in sensitive domains | Accuracy | Higher values denotes more knowledge in sensitive domains (cybersecurity, biosecurity, and chemical security) |
75
+
| Toxigen | Ability to detect toxic content | F1 Score | Higher values means better ability to detect toxic content |
76
+
77
+
### Model harmful behaviors
78
+
The [HarmBench](https://github.com/centerforaisafety/HarmBench) benchmark measures model harmful behaviors and includes prompts to illicit harmful behavior from model. As it relates to safety, the benchmark covers 7 semantic categories of behavior:
79
+
- Cybercrime & Unauthorized Intrusion
80
+
- Chemical & Biological Weapons/Drugs
81
+
- Copyright Violations
82
+
- Misinformation & Disinformation
83
+
- Harassment & Bullying
84
+
- Illegal Activities
85
+
- General Harm
86
+
87
+
These 7 categories can be summarized into 3 functional categories
88
+
- standard harmful behaviors
89
+
- contextually harmful behaviors
90
+
- copyright violations
91
+
92
+
Each functional category is featured in a separate scenario leaderboard. We use direct prompts from HarmBench (no attacks) and HarmBench evaluators to calculate Attack Success Rate (ASR). Lower ASR values means safer models. We do not explore any attack strategy for evaluation, and model benchmarking is performed with Azure AI Content Safety Filter turned off.
93
+
94
+
95
+
### Model ability to detect toxic content
96
+
[Toxigen](https://github.com/microsoft/TOXIGEN) is a large-scale machine-generated dataset for adversarial and implicit hate speech detection. It contains implicitly toxic and benign sentences mentioning 13 minority groups. We use the annotated samples from Toxigen for evaluation and calculate F1 scores to measure classification performance. Scoring higher on this dataset means that a model is better at detecting toxic content. Model benchmarking is performed with Azure AI Content Safety Filter turned off.
97
+
98
+
### Model knowledge in sensitive domains
99
+
The [Weapons of Mass Destruction Proxy](https://github.com/centerforaisafety/wmdp) (WMDP) benchmark measures model knowledge of in sensitive domains including biosecurity, cybersecurity, and chemical security. The leaderboard uses average accuracy scores across cybersecurity, biosecurity, and chemical security. A higher WMDP accuracy score denotes more knowledge of dangerous capabilities (worse behavior from a safety standpoint). Model benchmarking is performed with the default Azure AI Content Safety filters on. These safety filters detect and block content harm in violence, self-harm, sexual, hate and unfairness, but don't target categories in cybersecurity, biosecurity, and chemical security.
100
+
101
+
### Limitations of safety benchmarks
102
+
We understand and acknowledge that safety is a complex topic and has several dimensions. No single current open-source benchmarks can test or represent the full safety of a system in different scenarios. Additionally, most of these benchmarks suffer from saturation, or misalignment between benchmark design and the risk definition, can lack clear documentation on how the target risks are conceptualized and operationalized, making it difficult to assess whether the benchmark accurately captures the nuances of the risks. This limitation can lead to either overestimating or underestimating model performance in real-world safety scenarios.
103
+
65
104
## Performance benchmarks of language models
66
105
67
106
Performance metrics are calculated as an aggregate over 14 days, based on 24 trails (two requests per trail) sent daily with a one-hour interval between every trail. The following default parameters are used for each request to the model endpoint:
The following example demonstrates how to use the OpenAI SDK with Foundry Local. The code initializes the Foundry Local service, loads a model, and generates a response using the OpenAI SDK.
30
+
31
+
Copy-and-paste the following code into a C# file named `Program.cs`:
The following example demonstrates how to inference using a request to the Foundry Local service. The code initializes the Foundry Local service, loads a model, and generates a response using the `reqwest` library.
31
+
32
+
Copy-and-paste the following code into the Rust file named `main.rs`:
33
+
34
+
```rust
35
+
usefoundry_local::FoundryLocalManager;
36
+
useanyhow::Result;
37
+
38
+
#[tokio::main]
39
+
asyncfnmain() ->Result<()> {
40
+
// Create a FoundryLocalManager instance with default options
41
+
letmutmanager=FoundryLocalManager::builder()
42
+
.alias_or_model_id("qwen2.5-0.5b") // Specify the model to use
43
+
.bootstrap(true) // Start the service if not running
44
+
.build()
45
+
.await?;
46
+
47
+
// Use the OpenAI compatible API to interact with the model
0 commit comments