granite3-guardian

IBM Granite Guardian 3.0 **2B 和 8B 模型**旨在检测提示和/或响应中的风险。它们可以帮助检测 **IBM AI 风险图谱** 中列出的许多关键维度的风险。它们是在包含人工标注和基于内部红队攻击的合成数据的人类标注数据的独特数据上进行训练的，并且在标准基准测试中，它们优于同一领域的其他开源模型。

参数大小

该模型将生成单个输出标记，即 Yes 或 No。默认情况下，使用通用 harm 类别，但可以通过设置系统提示来选择其他类别。

ollama run granite3-guardian:2b
>>> /set system profanity

ollama run granite3-guardian:8b
>>> /set system violence

支持的使用场景

在提示文本或模型响应中检测风险（例如作为护栏），例如
- 有害（harm）：被认为普遍有害的内容
- 社会偏见（social_bias）：基于身份或特征的偏见
- 越狱（jailbreak）：故意操纵人工智能以生成有害、不希望或不适当的内容
- 暴力（violence）：宣扬身体、精神或性伤害的内容
- 亵渎（profanity）：使用攻击性语言或侮辱
- 色情内容（sexual_content）：具有性暗示或明显的性内容
- 不道德行为（unethical_behavior）：违反道德或法律标准的行为
RAG（检索增强生成）以评估
- 上下文相关性（relevance）：检索到的上下文是否与查询相关
- 扎根性（groundedness）：响应是否准确并忠实于提供的上下文
- 答案相关性（answer_relevance）：响应是否直接解决用户的查询

Granite 密集模型

Granite 密集模型提供 **2B 和 8B** 参数大小，旨在支持基于工具的使用案例以及检索增强生成 (RAG)，简化代码生成、翻译和错误修复。

查看模型页面

Granite 专家混合模型

Granite MoE 模型提供 **1B 和 3B** 参数大小，旨在用于低延迟使用，并支持部署在设备上的应用或需要即时推理的情况。

查看模型页面

了解更多

**开发者：**IBM Research
**GitHub 仓库：** ibm-granite/granite-guardian
**网站：** Granite Guardian 文档
**食谱：** Granite Guardian Snack
**发布日期：** 2024 年 10 月 21 日
**许可证：** Apache 2.0.

## Granite guardian models

The IBM Granite Guardian 3.0 **2B and 8B models** are designed to detect risks in prompts and/or responses. They can help with risk detection along many key dimensions catalogued in the [IBM AI Risk Atlas](https://www.ibm.com/docs/en/watsonx/saas?topic=ai-risk-atlas). They are trained on unique data comprising human annotations and synthetic data informed by internal red-teaming, and they outperform other open-source models in the same space on standard benchmarks.

### Parameter Sizes

The model will produce a single output token, either `Yes` or `No`. By default, the general-purpose `harm` category is used, but other categories can be selected by setting the system prompt.

**2B:**
  
```
ollama run granite3-guardian:2b
>>> /set system profanity
```

**8B:**

```
ollama run granite3-guardian:8b
>>> /set system violence
```

### Supported Uses

* Risk detection in prompt text or model response (i.e. as guardrails), such as:
    * Harm (`harm`): content considered generally harmful
    * Social Bias (`social_bias`): prejudice based on identity or characteristics
    * Jailbreaking (`jailbreak`): deliberate instances of manipulating AI to generate harmful, undesired, or inappropriate content
    * Violence (`violence`): content promoting physical, mental, or sexual harm
    * Profanity (`profanity`): use of offensive language or insults
    * Sexual Content (`sexual_content`): explicit or suggestive material of a sexual nature
    * Unethical Behavior (`unethical_behavior`): actions that violate moral or legal standards

* RAG (retrieval-augmented generation) to assess: 
    * Context relevance (`relevance`): whether the retrieved context is relevant to the query 
    * Groundedness (`groundedness`): whether the response is accurate and faithful to the provided context
    * Answer relevance (`answer_relevance`): whether the response directly addresses the user's query

## Granite dense models

The Granite dense models are available in **2B and 8B** parameter sizes designed to support tool-based use cases and for retrieval augmented generation (RAG), streamlining code generation, translation and bug fixing.

[See model page](https://ollama.org.cn/library/granite3-dense)

## Granite mixture of experts models

The Granite MoE models are available in **1B and 3B** parameter sizes designed for low latency usage and to support deployment in on-device applications or situations requiring instantaneous inference.

[See model page](https://ollama.org.cn/library/granite3-moe)

## Learn more

- **Developers:** IBM Research
- **GitHub Repository:** [ibm-granite/granite-guardian](https://github.com/ibm-granite/granite-guardian)
- **Website**: [Granite Guardian Docs](https://www.ibm.com/granite/docs/models/guardian/)
- **Cookbook**: [Granite Guardian Snack](https://github.com/ibm-granite-community/granite-snack-cookbook/blob/main/recipes/Granite_Guardian/Granite_Guardian_Detailed_Guide.ipynb)
- **Release Date**: October 21st, 2024
- **License:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0).

粘贴、拖放或点击上传图片（.png、.jpeg、.jpg、.svg、.gif）