On August 23, OpenAI announced the launch of GPT-3.5 Turbo-based fine-tuning features and updated APIs, allowing enterprises and developers to customize ChatGPT with their own data.
Fine-tuning is a method of training a specific model using an existing general-purpose language model (such as GPT-3.5). Although general-purpose language models have strong language understanding and generation capabilities, they are not domain-specific or task-specific. By fine-tuning and optimizing the general model on its own data, training a proprietary model can better adapt to specific usage scenarios. While retaining the powerful capabilities of the general-purpose language model, the accuracy and efficiency of the model are further improved.
This is equivalent to doing renovations on an already built house to make it more in line with your needs and preferences, rather than building a new house from scratch. As a result, a lot of time and resources can be saved, and some technical difficulties can also be avoided.
The fine-tuning function launched by OpenAI enables more developers to participate in GPT model applications, and uses this to achieve more personalized and innovative application scenarios, improve user experience and stickiness, and play a positive role in promoting the establishment of developer ecosystem. At the same time, it also greatly expands the application scope and potential of general models, and accelerates the pace of deploying AI technology in all walks of life.
The security, use effect, price, future updates, and deployment steps of GPT-3.5 Turbo mentioned in this announcement are highlighted below.
Security: Data sent from fine-tuning APIs is owned by the customer, and will not be used by OpenAI or any other organization to train models. At the same time, in order to ensure the security of model deployment, OpenAI detects harmful data that conflicts with security standards through the audit API and GPT-4-powered audit system. (Fine-tuning is good for providing users with more personalized service, but it also lowers the technical threshold, which may lead to irresponsible use)
Fine-tuned results: During actual testing, the fine-tuned version of the GPT-3.5 Turbo was comparable or even better than the GPT-4 capabilities of the base model on some tasks. Fine-tuning customers can improve model performance and reduce prompt time for common use cases. By fine-tuning the instructions of the model itself, API calls can also be accelerated and costs reduced, reducing the number of prompt words by 90%.
Price and Token: The fine-tuning cost of GPT-3.5 Turbo is divided into two parts: initial training cost and usage cost. A fine-tuning effort containing 100K tokens training files is expected to cost $2.40. Specifically, training: $0.008 / 1K tokens; Input used: $0.012 / 1K tokens; Output used: $0.016 / 1K tokens. The context that can be processed is 4K tokens, which is twice as many as the previous fine-tuned model.
Fine-tuning steps: You only need to go through four steps: preparing data, uploading files, creating fine-tuning jobs, and using fine-tuning models. Once the model has completed the fine-tuning process, it can be used immediately in production.
Future updates: Fine-tuning for GPT-4 will be available this fall, with fine-tuning support for GPT-3.5 function calls and 16k contexts coming later in the fall. A fine-tuning UI will also be available in the near future, making it easier to access information about ongoing fine-tuning jobs and more.
GPT-3 iterations: OpenAI is now offering babbage-002 and davinci-002 models as GPT-3 base models or fine-tuned models. The original GPT-3 base models (ADA, Babbage, Curie, Davinci) will be closed on January 4, 2024.
In summary, by opening up the fine-tuning capabilities of advanced large models, OpenAI can attract more enterprises and developers to use its platform and models, thereby expanding its influence and increasing revenue to alleviate loss pressure. This will help AI products be more widely used and accelerate the landing of large models.
At the same time, this feature update can also be seen as a response to Meta's open source model and allowing external competition such as commercial use. By allowing fine-tuning models, OpenAI can further expand and consolidate its user base and secure its leading position in the industry. This may also prompt other companies to accelerate the pace of commercialization, and further intensifies industry competition.
The following is the full text of the OpenAI announcement:
Fine-tuning for the GPT-3.5 Turbo is available now, and fine-tuning for GPT-4 will be available this fall. This update enables developers to customize models that better suit their use cases and run those custom models at scale. Early tests have shown that a fine-tuned version of the GPT-3.5 Turbo can match or even better than the underlying GPT-4 in certain vertical tasks. As with all of our APIs, the data sent from the fine-tuning API is owned by the customer and is not used by OpenAI or any other organization to train other models.
Fine-tune use cases
Since the release of GPT-3.5 Turbo, developers and enterprises have demanded the ability to customize models to create unique and differentiated experiences for their users. With this launch, developers can now run supervised fine-tuning to make the model perform better in their use cases.
In our private beta, fine-tuning customers have been able to effectively improve model performance for common use cases, such as:
Improved manipulability: Fine-tuning allows businesses to make models better at following instructions, such as making output concise or always responding in a given language. For example, developers can use fine-tuning to ensure that models always respond as required in German.
Reliable output format: Fine-tuning improves the model's ability to format responses consistently, which is critical for applications that require a specific response format, such as code completion or composing API calls. Developers can use fine-tuning to more reliably translate user prompts into high-quality JSON snippets that can be used with their own systems.
Custom style: Fine-tuning is a great way to hone the qualitative feel of the model's output, and businesses with a recognizable brand style can use fine-tuning to make the model more consistent with its tone.
In addition to improving performance, fine-tuning enables businesses to reduce prompt times while ensuring similar performance. Fine-tuning with GPT-3.5-Turbo can also handle 4K tokens – twice as many as our previous fine-tuning model. Early testers reduced prompt size by up to 90% by fine-tuning instructions to the model itself, speeding up each API call and reducing costs.
Fine-tuning is most powerful when combined with other techniques such as prompt engineering, information retrieval, and function calls. To learn more about this, check out our fine-tuning guide. In addition, fine-tuning support for function calls and GPT-3.5-turbo-16k will take place later this fall.
Fine-tune the steps
We'll also be rolling out a fine-tuning UI in the near future, which will make it easier for developers to access information about ongoing fine-tuning jobs, completed model snapshots, and more.
safe
It is very important for us to deploy fine-tuning securely. In order to preserve the security features of the default model during fine-tuning, fine-tuning training data will go through our audit API and GPT-4 supported auditing system to detect insecure training data that conflicts with our security standards.
Pricing
Fine-tuning costs fall into two categories: initial training costs and usage costs. Training: $0.008 / 1K tokens; Input used: $0.012 / 1K tokens; Output used: $0.016 / 1K tokens. For example, fine-tuning work with 100,000 tokens is expected to cost $2.40.
Updated GPT-3 model
In July of this year, we announced that the original GPT-3 base model would be closed on January 4, 2024. Today we are manufacturing alternatives to these models, both as a base model and as a fine-tuning model. These models can be fine-tuned using our new API endpoints. Pricing for basic and fine-tuned GPT-3 models is as follows:
The new endpoints provide paging and more extensibility to support future fine-tuning APIs, and transitioning to newer endpoints is easy, as detailed in our fine-tuning guide.