I am currently fine-tuning the GPT-3.5 Turbo model with my custom dataset using OpenAI’s dashboard, and I am able to achieve the desired results on the Playground. However, I would like to use this fine-tuned model as an API in another application. My question is whether it needs to be deployed externally, or if it is automatically saved within OpenAI’s environment.
Additionally, are there any costs beyond token usage, such as storage or hosting fees for the model saved on OpenAI’s servers? If so, how can I determine these costs?
There is no separate deployment required. When specifying the model in your API request, simply reference the fine-tuned model’s name just like you would with a regular model.
Fine-tuned models created with OpenAI do not incur any storage or hosting fees. This is different from deploying a fine-tuned model with Azure where these costs occur.
Thank you for the clarification. Your response addressed my question. However, if I train my model using OpenAI’s dashboard, my dataset needs to be in a JSONL format. Currently, my dataset is in PDF or DOC format, containing segments or paragraphs, such as organizational policies or legal terms and conditions. In this case, what steps should I take, and do you have any suggestions for how to proceed?
It sounds to me like you are looking to build a solution that allows answering of questions on the basis of your proprietary materials.
For this case, fine-tuning is actually not suitable. Instead you should look into RAG solutions. Fine-tuning is intended to get the model to respond in a certain style or format, or to get the model to execute a task in a specific way. Information / knowledge contained in fine-tuning datasets is not systematically retained. You can read up more in this distinction in in this OpenAI guide.
In the above scenario, if we choose a RAG solution, could someone recommend the most secure platform for deploying my code? and what are the steps to implement it ?