mannix / phi3-mini-4k

Phi-3 Mini 是微软的先进开放式 3B 模型，轻量级。2024年7月更新。

139 pulls Updated 7 weeks ago

Updated 7 weeks ago

7 weeks ago

c01f1d317007 · 2.2GB

Readme

从 fp32 进行量化
使用 i-matrix 校准数据 calibration_datav3.txt

Phi-3 是微软开发的一系列开源 AI 模型。

参数大小

Phi-3 Mini – 3B 参数

上下文窗口大小

4k ollama run mannix/phi3-mini-4k

Phi-3 Mini

Phi-3 Mini 是一个 3.8B 参数，轻量级，最先进的开源模型，使用 Phi-3 数据集经过训练，其中包括合成的数据以及过滤的公开网站数据，重点在于高质量和密集推理属性。

该模型经过了一项训练后处理流程，结合了监督微调和直接偏好优化，以确保精确遵循指令并采取稳健的安全措施。

当与测试常识、语言理解、数学、代码、长语境和逻辑推理的基准测试进行评估时，Phi-3 Mini-4K-Instruct在参数少于13亿的模型中展示出稳健和最先进的表现。

Phi-3 Medium

Phi-3 Medium是一个14B参数的语言模型，其表现优于Gemini 1.0 Pro。

预期用途

主要用例

该模型旨在为英语的商业和研究用途提供服务。该模型适用于需要以下功能的应用程序
1) 存储和计算受限的环境
2) 延迟限制场景
3) 强推理（尤其是数学和逻辑）
4) 长语境

我们的模型旨在加速语言和多媒体模型的研究，作为生成AI驱动功能的构建块。

用例考虑因素

我们的模型并非专门设计或评估用于所有下游用途。开发者在选择用例时，应考虑语言模型的常见局限性，并在特定下游用例中使用之前，评估和缓解准确性、安全性和公平性，尤其是在高风险场景中。
开发者应了解并遵守与他们用例相关的适用法律或法规（包括隐私、贸易合规法规等）。
本模型卡中包含的内容不应被解释为或视为对模型发布许可证的限制或修改。

负责任的AI考虑因素

与其他语言模型一样，Phi系列模型可能以不公平、不可靠或冒犯性的方式行事。应了解的某些限制行为包括

服务质量：Phi模型主要训练于英语文本。除英语以外的语言将体验到更差的表现。在训练数据中代表性较低的非标准美国英语可能会比标准美国英语体验更差的表现。
危害表示及刻板印象的延续：这些模型可能过度或不足地表示某些人群，消除某些人群的表示，或加强贬低或负面的刻板印象。尽管经过安全训练，但由于不同群体的表示水平或训练数据中负面刻板印象的普遍性存在差异，这些局限性仍然可能存在。
不适或冒犯性内容：这些模型可能产生其他类型的不适或冒犯性内容，在没有针对特定用例采取额外缓解措施的情况下，可能不适合在敏感情境下部署。
信息可靠性：语言模型可以生成无意义的内容或捏造看似合理但实际上不准确或过时的内容。
代码范围有限：Phi-3大部分训练数据基于Python，使用常见的包，如“typing，math，random，collections，datetime，itertools”。如果模型生成利用其他包或使用其他语言的Python脚本，我们强烈建议用户手动验证所有API的使用。

开发者应应用负责任的AI最佳实践，并负责确保特定用例遵守相关法律和法规（例如：隐私、贸易等）。重要的考虑领域包括
+ 分配：在没有进一步评估和额外的去偏斜技术的情况下，模型可能不适合对法律地位、资源分配或生活机会有重大影响的场景（例如：住房、就业、信贷等）

高风险场景：开发者在使用模型面临不公平、不可靠或冒犯性的输出可能极其昂贵或导致伤害的高风险场景时，应评估其适用性。这包括在敏感或专业领域提供建议，其中准确性和可靠性至关重要（例如，法律或健康建议）。应根据部署环境在应用级别实施额外的安全措施。
虚假信息：模型可能产生不准确的信息。开发者应遵循透明度最佳实践，并告知最终用户他们正在与一个AI系统交互。在应用级别，开发者可以构建反馈机制和管道，以将响应建立在特定用例的上下文信息基础上，这是一种称为检索增强生成（RAG）的技术。
生成有害内容：开发者应评估输出在特定情境和用途下的适用性，并使用适用于他们用例的安全分类器或定制解决方案。
滥用：其他形式的滥用，如欺诈、垃圾邮件或恶意软件制作，也可能发生，开发者应确保他们的应用程序不违反适用的法律和法规。

训练

模型

架构：Phi-3 Mini具有3.8B个参数，是一个密集型仅解码Transformer模型。该模型通过监督微调（SFT）和直接偏好优化（DPO）进行微调，以确保与人类偏好和安全指南保持一致。
输入：文本。它最适合使用聊天格式进行提示。
上下文长度：128K个令牌
GPU：512个H100-80G
训练时间：7天
训练数据：3.3T个令牌
输出：对输入生成的文本
日期：我们的模型训练时间在2024年2月至4月之间
状态：这是一个在截至2023年10月的离线数据集上训练的静态模型。随着我们改进模型，未来版本的微调模型可能会发布。

数据集

我们的训练数据包括来源广泛的文档，共计3300亿个令牌，并包括以下内容：
1）经过严格质量筛选的公开可用的文档，选用了高质量的学术数据和教育数据，以及代码；
2）为新教学目的而创建的合成“教科书级”数据，用于教授数学、编码、常识推理、世界一般知识（科学、日常活动、心智理论等）；
3）涵盖各种主题的高质量聊天格式监督数据，以反映人类在不同方面的偏好，例如指令遵循、真实性、诚实和有用性。

软件

许可

该模型遵循MIT许可证。

商标

本项目可能包含项目、产品或服务的商标或标志。未经授权使用Microsoft商标或标志需遵守和遵循Microsoft的商标和品牌指南。修改本项目时使用Microsoft商标或标志不得引起混淆或暗示得到微软的支持。对第三方商标或标志的使用受第三方政策约束。

资源

![image.png](https://ollama.org.cn/assets/mannix/phi3-mini-4k/b030ede8-9cbe-4d5c-97d2-6446495c2214)

- Quantization from `fp32`
- Using i-matrix `calibration_datav3.txt`

<p>Phi-3 is a family of open AI models developed by Microsoft.</p>

<h2>Parameter sizes</h2>

<ul>
<li><a href="https://ollama.org.cn/library/phi3:mini" rel="nofollow">Phi-3 Mini</a> – 3B parameters<br>
</li>
</ul>

<h2>Context window sizes</h2>

<ul>
<li>4k <code>ollama run mannix/phi3-mini-4k</code><br>
</li>
</ul>

<p>Phi-3 Mini is a 3.8B parameters, lightweight, state-of-the-art open model trained with the Phi-3 datasets that includes both synthetic data and the filtered publicly available websites data with a focus on high-quality and reasoning dense properties.</p>

<p>The model has underwent a post-training process that incorporates both supervised fine-tuning and direct preference optimization to ensure precise instruction adherence and robust safety measures.</p>

<p>When assessed against benchmarks testing common sense, language understanding, math, code, long context and logical reasoning, Phi-3 Mini-4K-Instruct showcased a robust and state-of-the-art performance among models with less than 13 billion parameters.</p>

<h2>Phi-3 Medium</h2>

<p>Phi-3 Medium is a 14B parameter language model, and outperforms Gemini 1.0 Pro.</p>

<h2>Intended Uses</h2>

<p><strong>Primary use cases</strong></p>

<p>The model is intended for commercial and research use in English. The model provides uses for applications which require<br>
1) memory/compute constrained environments<br>
2) latency bound scenarios<br>
3) strong reasoning (especially math and logic)<br>
4) long context</p>

<p>Our model is designed to accelerate research on language and multimodal models, for use as a building block for generative AI powered features.</p>

<p><strong>Use case considerations</strong></p>

<p>Our models are not specifically designed or evaluated for all downstream purposes. Developers should consider common limitations of language models as they select use cases, and evaluate and mitigate for accuracy, safety, and fariness before using within a specific downstream use case, particularly for high risk scenarios.<br>
Developers  should be aware of and adhere to applicable laws or regulations (including privacy, trade compliance laws, etc.) that are relevant to their use case.<br>
Nothing contained in this Model Card should be interpreted as or deemed a restriction or modification to the license the model is released under.</p>

<h2>Responsible AI Considerations</h2>

<p>Like other language models, the Phi series models can potentially behave in ways that are unfair, unreliable, or offensive. Some of the limiting behaviors to be aware of include:</p>

<ul>
<li><p>Quality of Service: the Phi models are trained primarily on English text. Languages other than English will experience worse performance. English language varieties with less representation in the training data might experience worse performance than standard American English.</p></li>

<li><p>Representation of Harms &amp; Perpetuation of Stereotypes: These models can over- or under-represent groups of people, erase representation of some groups, or reinforce demeaning or negative stereotypes. Despite safety post-training, these limitations may still be present due to differing levels of representation of different groups or prevalence of examples of negative stereotypes in training data that reflect real-world patterns and societal biases.</p></li>

<li><p>Inappropriate or Offensive Content: these models may produce other types of inappropriate or offensive content, which may make it inappropriate to deploy for sensitive contexts without additional mitigations that are specific to the use case.</p></li>

<li><p>Information Reliability: Language models can generate nonsensical content or fabricate content that might sound reasonable but is inaccurate or outdated.</p></li>

<li><p>Limited Scope for Code: Majority of Phi-3 training data is based in Python and use common packages such as “typing, math, random, collections, datetime, itertools”. If the model generates Python scripts that utilize other packages or scripts in other languages, we strongly recommend users manually verify all API uses.</p></li>
</ul>

<p>Developers should apply responsible AI best practices and are responsible for ensuring that a specific use case complies with relevant laws and regulations (e.g. privacy, trade, etc.). Important areas for consideration include:<br>
+ Allocation: Models may not be suitable for scenarios that could have consequential impact on legal status or the allocation of resources or life opportunities (ex: housing, employment, credit, etc.) without further assessments and additional debiasing techniques.</p>

<ul>
<li><p>High-Risk Scenarios: Developers should assess suitability of using models in high-risk scenarios where unfair, unreliable or offensive outputs might be extremely costly or lead to harm. This includes providing advice in sensitive or expert domains where accuracy and reliability are critical (ex: legal or health advice). Additional safeguards should be implemented at the application level according to the deployment context.</p></li>

<li><p>Misinformation: Models may produce inaccurate information. Developers should follow transparency best practices and inform end-users they are interacting with an AI system. At the application level, developers can build feedback mechanisms and pipelines to ground responses in use-case specific, contextual information, a technique known as Retrieval Augmented Generation (RAG).</p></li>

<li><p>Generation of Harmful Content: Developers should assess outputs for their context and use available safety classifiers or custom solutions appropriate for their use case.</p></li>

<li><p>Misuse: Other forms of misuse such as fraud, spam, or malware production may be possible, and developers should ensure that their applications do not violate applicable laws and regulations.</p></li>
</ul>

<h2>Training</h2>

<h3>Model</h3>

<ul>
<li>Architecture: Phi-3 Mini has 3.8B parameters and is a dense decoder-only Transformer model. The model is fine-tuned with Supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) to ensure alignment with human preferences and safety guidelines.<br>
</li>
<li>Inputs: Text. It is best suited for prompts using chat format.<br>
</li>
<li>Context length: 128K tokens<br>
</li>
<li>GPUS: 512 H100-80G<br>
</li>
<li>Training time: 7 days<br>
</li>
<li>Training data: 3.3T tokens<br>
</li>
<li>Outputs: Generated text in response to the input<br>
</li>
<li>Dates: Our models were trained between February and April 2024<br>
</li>
<li>Status: This is a static model trained on an offline dataset with cutoff date October 2023. Future versions of the tuned models may be released as we improve models.<br>
</li>
</ul>

<h3>Datasets</h3>

<p>Our training data includes a wide variety of sources, totaling 3.3 trillion tokens, and is a combination of<br>
1) publicly available documents filtered rigorously for quality, selected high-quality educational data, and code;<br>
2) newly created synthetic, “textbook-like” data for the purpose of teaching math, coding, common sense reasoning, general knowledge of the world (science, daily activities, theory of mind, etc.);<br>
3) high quality chat format supervised data covering various topics to reflect human preferences on different aspects such as instruct-following, truthfulness, honesty and helpfulness.</p>

<h3>Software</h3>

<ul>
<li><a href="https://github.com/pytorch/pytorch" rel="nofollow">PyTorch</a><br>
</li>
<li><a href="https://github.com/microsoft/DeepSpeed" rel="nofollow">DeepSpeed</a><br>
</li>
<li><a href="https://github.com/huggingface/transformers" rel="nofollow">Transformers</a><br>
</li>
<li><a href="https://github.com/HazyResearch/flash-attention" rel="nofollow">Flash-Attention</a><br>
</li>
</ul>

<h3>License</h3>

<p>The model is licensed under the <a href="https://ollama.org.cn/library/phi3:latest/blobs/fa8235e5b48fENSE" rel="nofollow">MIT license</a>.</p>

<h2>Trademarks</h2>

<p>This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow <a href="https://www.microsoft.com/en-us/legal/intellectualproperty/trademarks" rel="nofollow">Microsoft’s Trademark &amp; Brand Guidelines</a>. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party’s policies.</p>

<h2>Resources</h2>

<ul>
<li><a href="https://hugging-face.cn/microsoft/Phi-3-mini-4k-instruct-gguf" rel="nofollow">HuggingFace</a><br>
</li>
<li><a href="https://aka.ms/phi3-blog" rel="nofollow">Phi-3 Microsoft Blog</a><br>
</li>
<li><a href="https://aka.ms/phi3-tech-report" rel="nofollow">Phi-3 Technical Report</a><br>
</li>
<li><a href="https://aka.ms/phi3-azure-ai" rel="nofollow">Phi-3 on Azure AI Studio</a><br>
</li>
<li><a href="https://aka.ms/phi3-hf" rel="nofollow">Phi-3 on Hugging Face</a><br>
</li>
<li>Phi-3 ONNX: <a href="https://aka.ms/phi3-mini-4k-instruct-onnx" rel="nofollow">4K</a> and <a href="https://aka.ms/phi3-mini-128k-instruct-onnx" rel="nofollow">128K</a></li>
</ul>

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)