[模型微调] What is the difference between ChatGPT's GPS and its large-model fine-tuning, and what scenarios are they respectively suitable for?

GPTs are not equivalent to fine-tuned large models; the gap is huge. Although both are trained in the form of a knowledge base and output content, there are huge differences in support and functionality. Now we will introduce the differences between the two in detail.

Differences in Functionality and Support

1. GPTs have limited upload capacity; this feature supports only 10 files and a single file with a character data volume within 8000 KB. The fine-tuned large model strategy is different: there is almost no limit on data volume, but the more training samples, the higher the charge.
2. As a free adjunct feature of ChatGPT-4, GPTs' token usage limit is included in ChatGPT-4, so the 40 sessions every 3 hours are included; in addition, for undisclosed reasons, other usage-occupying reasons further limit usage. Fine-tuned large models have no such limit, depending on your API plan level.
3. GPTs have extremely poor response speed; each session feedback requires first querying the requirements and knowledge base files and API, and must be recomputed every time, so they cannot answer immediately. A fine-tuned large model, as the default model, responds immediately.
4. GPTs are harder to use, very easy to confuse, and prone to rollback. Once a large model is well trained, problems almost never occur.
5. GPTs can only process text and images, and their functionality appears weak. Large models can be trained on text, images, video, and audio.


Applicable Scenarios


GPTs are suitable for very small, single-function domains and for samples with low computational demand; however, as an adjunct for Plus members, being able to experience the large-model training process is already quite good.
Large models are suitable for complex and precise functions and require a longer period of tuning; once well tuned, they become a powerful enterprise tool.