2024年4月4日，OpenAI发文宣布推出新的微调 API 功能。 在微调 API 中引入了新的仪表板、指标和集成，为开发人员提供更多控制权，并添加了使用 OpenAI 构建自定义模型的新方法。

▌官方博客原文（含中英文）

Introducing improvements to the fine-tuning API and expanding our custom models program

引入微调 API 的改进并扩展我们的自定义模型计划

We’re adding new features to help developers have more control over fine-tuning and announcing new ways to build custom models with OpenAI.

我们正在添加新功能，以帮助开发人员更好地控制微调，并宣布使用 OpenAI 构建自定义模型的新方法。

There are a variety of techniques that developers can use to increase model performance in an effort to reduce latency, improve accuracy, and reduce costs. Whether it’s extending model knowledge with retrieval-augmented generation (RAG), customizing a model’s behavior with fine-tuning, or building a custom-trained model with new domain-specific knowledge, we have developed a range of options to support our customers’ AI implementations. Today, we’re launching new features to give developers more control over fine-tuning with the API and introducing more ways to work with our team of AI experts and researchers to build custom models.

开发人员可以使用多种技术来提高模型性能，以减少延迟、提高准确性并降低成本。无论是通过检索增强生成 (RAG) 扩展模型知识、通过微调自定义模型的行为，还是使用新的特定领域知识构建自定义训练的模型，我们都开发了一系列选项来支持客户的 AI实施。今天，我们推出新功能，让开发人员能够更好地控制 API 的微调，并引入更多方法与我们的 AI 专家和研究人员团队合作构建自定义模型。

New fine-tuning API features

新的微调 API 功能

We launched the self-serve fine-tuning API for GPT-3.5 in August 2023. Since then, thousands of organizations have trained hundreds of thousands of models using our API. Fine-tuning can help models deeply understand content and augment a model’s existing knowledge and capabilities for a specific task. Our fine-tuning API also supports a larger volume of examples than can fit in a single prompt to achieve higher quality results while reducing cost and latency. Some of the common use cases of fine-tuning include training a model to generate better code in a particular programming language, to summarize text in a specific format, or to craft personalized content based on user behavior.

我们于 2023 年 8 月推出了 GPT-3.5 的自助微调 API。从那时起，数千个组织已经使用我们的 API 训练了数十万个模型。微调可以帮助模型深入理解内容并增强模型针对特定任务的现有知识和能力。我们的微调 API 还支持比单个提示所能容纳的更多示例，以实现更高质量的 结果，同时降低成本和延迟。微调的一些常见用例包括训练模型以特定编程语言生成更好的代码、以特定格式总结文本或根据用户行为制作个性化内容。

For example, Indeed, a global job matching and hiring platform, wants to simplify the hiring process. As part of this, Indeed launched a feature that sends personalized recommendations to job seekers, highlighting relevant jobs based on their skills, experience, and preferences. They fine-tuned GPT-3.5 Turbo to generate higher quality and more accurate explanations. As a result, Indeed was able to improve cost and latency by reducing the number of tokens in prompt by 80%. This let them scale from less than one million messages to job seekers per month to roughly 20 million.

例如， 全球职位匹配和招聘平台Indeed希望简化招聘流程。作为其中的一部分，Indeed 推出了一项功能，向求职者发送个性化推荐，根据求职者的技能、经验和偏好突出显示相关工作。他们对 GPT-3.5 Turbo 进行了微调，以生成更高质量和更准确的解释。因此，Indeed 能够通过将提示中的代币数量减少 80% 来改善成本和延迟。这使得他们每月向求职者发送的消息数量从不到 100 万条增加到大约 2000 万条。

Today, we’re introducing new features to give developers even more control over their fine-tuning jobs, including:

• Epoch-based Checkpoint Creation: Automatically produce one full fine-tuned model checkpoint during each training epoch, which reduces the need for subsequent retraining, especially in the cases of overfitting

• Comparative Playground : A new side-by-side Playground UI for comparing model quality and performance, allowing human evaluation of the outputs of multiple models or fine-tune snapshots against a single prompt

• Third-party Integration: Support for integrations with third-party platforms (starting with Weights and Biases this week) to let developers share detailed fine-tuning data to the rest of their stack

• Comprehensive Validation Metrics : The ability to compute metrics like loss and accuracy over the entire validation dataset instead of a sampled batch, providing better insight on model quality

• Hyperparameter Configuration : The ability to configure available hyperparameters from the Dashboard (rather than only through the API or SDK)

• Fine-Tuning Dashboard Improvements : Including the ability to configure hyperparameters, view more detailed training metrics, and rerun jobs from previous configurations

• 基于 Epoch 的检查点创建：在每个训练 epoch 期间自动生成一个完整的微调模型检查点，这减少了后续重新训练的需要，尤其是在过度拟合的情况下
• Comparison Playground：一个新的并排 Playground UI，用于比较模型质量和性能，允许对多个模型的输出进行人工评估或根据单个提示微调快照
• 第三方集成：支持与第三方平台集成（从本周的 权重和偏差开始），让开发人员与堆栈的其余部分共享详细的微调数据
• 全面的验证 指标：能够计算整个验证数据集（而不是采样批次）的损失和准确性等指标，从而更好地了解模型质量
• 超参数配置：能够从 仪表板配置可用的超参数（而不仅仅是通过 API 或 SDK）
• 微调仪表板改进：包括配置超参数、查看更详细的训练指标以及从以前的配置重新运行作业的能力

Expanding our Custom Models Program

扩大我们的定制模型计划

At DevDay last November, we announced a Custom Model program designed to train and optimize models for a specific domain, in partnership with a dedicated group of OpenAI researchers. Since then, we've met with dozens of customers to assess their custom model needs and evolved our program to further maximize performance.

在去年 11 月的 DevDay 上，我们宣布了一项自定义模型计划，旨在与专门的 OpenAI 研究人员小组合作，训练和优化特定领域的模型。从那时起，我们与数十位客户会面，评估他们的定制模型需求 ，并改进我们的计划以进一步最大限度地提高性能。

Today, we are formally announcing our assisted fine-tuning offering as part of the Custom Model program. Assisted fine-tuning is a collaborative effort with our technical teams to leverage techniques beyond the fine-tuning API, such as additional hyperparameters and various parameter efficient fine-tuning (PEFT) methods at a larger scale. It’s particularly helpful for organizations that need support setting up efficient training data pipelines, evaluation systems, and bespoke parameters and methods to maximize model performance for their use case or task.

今天，我们正式宣布我们的辅助微调产品作为定制模型计划的一部分。辅助微调是我们与技术团队合作的成果，旨在利用微调 API 之外的技术，例如更大规模的额外超参数和各种参数高效微调 (PEFT) 方法。对于需要支持建立高效的训练数据管道、评估系统以及定制参数和方法以最大限度地提高其用例或任务的模型性能的组织来说，它特别有帮助。

For example, SK Telecom, a telecommunications operator serving over 30 million subscribers in South Korea, wanted to customize a model to be an expert in the telecommunications domain with an initial focus on customer service. They worked with OpenAI to fine-tune GPT-4 to improve its performance in telecom-related conversations in the Korean language. Over the course of multiple weeks, SKT and OpenAI drove meaningful performance improvement in telecom customer service tasks—a 35% increase in conversation summarization quality, a 33% increase in intent recognition accuracy, and an increase in satisfaction scores from 3.6 to 4.5 (out of 5) when comparing the fine-tuned model to GPT-4.

例如，SK Telecom是一家为韩国超过 3000 万用户提供服务的电信运营商，希望定制一个模型，成为电信领域的专家，最初的重点是客户服务。他们与 OpenAI 合作对 GPT-4 进行微调，以提高其在韩语电信相关对话中的性能。在数周的时间里，SKT 和 OpenAI 推动了电信客户服务任务的显着性能改进——对话摘要质量提高了 35%，意图识别准确性提高了 33%，满意度得分从 3.6 提高到 4.5（超出5）将微调模型与 GPT-4 进行比较。

In some cases, organizations need to train a purpose-built model from scratch that understands their business, industry, or domain. Fully custom-trained models imbue new knowledge from a specific domain by modifying key steps of the model training process using novel mid-training and post-training techniques. Organizations that see success with a fully custom-trained model often have large quantities of proprietary data—millions of examples or billions of tokens—that they want to use to teach the model new knowledge or complex, unique behaviors for highly specific use cases.

在某些情况下，组织需要从头开始训练一个专门构建的模型，以了解其业务、行业或领域。完全定制训练的模型通过使用新颖的训练中期和训练后技术修 改模型训练过程的关键步骤，注入特定领域的新知识。通过完全定制训练的模型取得成功的组织通常拥有大量专有数据（数百万个示例或数十亿个tokens），他们希望使用这些 数据来向模型传授新知识或针对高度特定的用例的复杂、独特的行为。

For example, Harvey, an AI-native legal tool for attorneys, partnered with OpenAI to create a custom-trained large language model for case law. While foundation models were strong at reasoning, they lacked the extensive knowledge of legal case history and other knowledge required for legal work. After testing out prompt engineering, RAG, and fine-tuning, Harvey worked with our team to add the depth of context needed to the model—the equivalent of 10 billion tokens worth of data. Our team modified every step of the model training process, from domain-specific mid-training to customizing post-training processes and incorporating expert attorney feedback. The resulting model achieved an 83% increase in factual responses and attorneys preferred the customized model’s outputs 97% of the time over GPT-4.

例如，Harvey是一款面向律师的 AI 原生法律工具，它与 OpenAI 合作，为判例法创建了定制训练的大型语言模型。虽然基础模型的推理能力很强，但它们缺乏广泛的法律案例历史知识和法律工作所需的其他知识。在测试了即时工程、RAG 和微调后，Harvey 与我们的团队合作，添加了模型所需的上下文深度——相当于 100 亿tokens的数据。我们的团队修改了模型培训过程的每一步，从特定领域的中期培训到定制培训后流程并纳入专家律师的反馈。由此产生的模型的事实答复增加了 83%，并且与 GPT-4 相比，律师在 97% 的时间里更喜欢定制模型的输出。

What’s next for model customization

模型定制的下一步是什么

We believe that in the future, the vast majority of organizations will develop customized models that are personalized to their industry, business, or use case. With a variety of techniques available to build a custom model, organizations of all sizes can develop personalized models to realize more meaningful, specific impact from their AI implementations. The key is to clearly scope the use case, design and implement evaluation systems, choose the right techniques, and be prepared to iterate over time for the model to reach optimal performance.

我们相信，未来绝大多数组织将开发针对其行业、业务或用例的个性化定制模型。借助多种可用于构建自定义模型的技术，各种规模的组织都可以开发个 性化模型，以从其人工智能实施中实现更有意义、更具体的影响。关键是要明确用例范围，设计和实施评估系统，选择正确的技术，并准备好随着时间的推移进行迭代以使模型达到最 佳性能。

With OpenAI, most organizations can see meaningful results quickly with the self-serve fine-tuning API. For any organizations that need to more deeply fine-tune their models or imbue new, domain-specific knowledge into the model, our Custom Model programs can help.

借助 OpenAI，大多数组织都可以通过自助微调 API 快速看到有意义的结果。对于任何需要更深入地调整模型或将新的、特定领域的知识注入模型的组织，我们的自定义模型计划可以提供帮助。

Visit our fine-tuning API docs to start fine-tuning our models. For more information on how we can help customize models for your use case, reach out to us.

请访问我们的微调 API文档以开始微调我们的模型。有关我们如何帮助您定制适用于您的用例的模型的更多信息，请联系我们。

▌ 内容来源

[1] OpenAI 官方博客 https://openai.com/blog/introducing-improvements-to-the-fine-tuning-api-and- expanding-our-custom-models-program

[2] OpenAI X平台账号发文