[模型微调] Role-playing, simulating characters, and training a unique copywriting model to output content.

A female influencer abroad used data from training her own Twitter social media account to create a GPT version of herself to chat with her fans, but the chat is paid. With the support of her devoted fans, she earns tens of thousands of dollars per month.

[baike]Role-playing scenarios[/baike]ChatGPT can play roles and is very useful in certain scenarios, such as brainstorming, roundtable gatherings, and co-creation sessions. You can simulate Elon Musk, Jack Ma, and Richard Liu discussing the future of the internet with you, and you can also have them guide you on how to improve at work. You can also simulate different types of people. Suppose you are a salesperson; you can ask ChatGPT to simulate your boss, your customer, your sales support, and your product manager to provide different suggestions for your plan to improve sales rates. You can imitate economic gurus and real estate tycoons to discuss future housing prices and help you judge whether now is a suitable time to invest in real estate.

You can also use it in reverse. For example, if you are a leader and have a new employee under you, you can ask ChatGPT to imitate her, the problems she may face, how to guide her to improve, how the two of you should cooperate, what your goals are, and how to achieve them—all of these can bring you substantive help.

[baike]Four-step training method[/baike]We divide the work of training a model into four steps. Do these four steps well, and you can have a unique content-output bot.

1. Choose the target person
First, determine the person you want to simulate. This can be a historical figure, a celebrity, or a fictional, living, or deceased person. But the most important thing is to have an in-depth understanding of this person, including their speech and behavior, writing style, values, etc.

2. Collect training data
Collect as much textual material about the target person as possible. This includes their speeches, articles, letters, social media posts, etc. These texts will serve as the basis for ChatGPT training.

The simplest way to collect content is to copy and paste, but this method is too inefficient. This is where two essential tools in AI work come in: a universal downloader and a web crawler.
Universal downloader: https://www.niuboyi.com/thread-634-1-1.html
Web crawler: https://www.niuboyi.com/thread-631-1-1.html

3. Upload training data

Before uploading the training data, we need to consolidate the data. For example, if the content is in multiple files, you can concatenate these files together, which makes it easier to categorize and upload. The ordinary ChatGPT-4 chat window only supports uploading one file at a time, so consolidation is necessary. Of course, I recommend creating a custom GPT, which supports batch uploads and can be set once for lifetime use.

4. Set the output format
A complete content structure should include: content style, content structure, writing method, tone and emotion. Uploading materials about the person you want to simulate does not mean you can fully simulate them. You need appropriate guidance and continuous optimization to get the desired effect. The initial prompt is as follows:

Learn the content of the uploaded document, fully understand the author's [content style, content structure, writing method, tone and emotion], and simulate the author's output according to the user's requirements.

For a better understanding, I suggest that after uploading, you consult ChatGPT to analyze the uploaded document and what you understand about the author's content. Then you can ask it what the [content style, content structure, writing method, tone and emotion] are like. Then ask, can you imitate the author to output content?