For several weeks, the paid version of ChatGPT has offered an option to create an artificial intelligence tool adapted to your specific needs.
If you have subscribed to the paid option of ChatGPT, OpenAI’s artificial intelligence, a new tab has appeared for some time at the top of the left column. “Explore” gives access to personalized formats (in data analysis, to explain the rules of board games, to learn the art of negotiation, etc.). But the most striking new feature is that you can create your own version of ChatGPT, personalized according to your needs.
What can a customized version of GPT4 do? It can respond in a specific tone, in a specific format, base its responses on data that you select and provide, help you to be more creative or even apply a process in the form of lines of code.
Concretely, this can be useful for the creation of newsletters, customer relationship management or applications in specific fields such as medicine, law, etc. By providing certain targeted information and data whose quality you guarantee, you ensure that the responses generated will be reliable. Better: as you use it and specify your needs, your GPT will become more and more
effective. Let’s take a look at the steps to establishing your first custom GPT.
1: Define your needs
There is no point in going head to head, or there is a good chance you will be disappointed. For your first GPT, find a concrete objective, not too complex. This can be for a specific domain (such as health, finance, education, etc.), for a particular type of task (generating content for a social network, creating code), etc. Next, determine specific needs in terms of tone, style and format of
responses. Response in French or English? Tone informative or relaxed? Length of expected
answers, formatting, etc.
To avoid getting too lost, it is better to stick to a GPT with a simple and concrete objective.
2: Define the data used
If necessary, gather a set of data relevant to your field or goal. This data may include texts, conversations, articles, manuals, etc. They must be representative of the type of content that the model will need to generate or understand. For example, when creating a newsletter with a certain tone and formatting, do not hesitate to create several versions that meet your
expectations.
This is not wasted time, on the contrary, it will serve as a guideline for your GPT. You may also need to clean and prepare this data for training, which involves structuring it coherently and ridding it of unwanted or irrelevant elements. So, in an Excel file, delete unnecessary columns to keep only what is relevant to your objective. You can then upload these documents to GPT.
3: Train Him
Like Rocky who will train rigorously, improving his strength, speed and technique to prepare specifically for the opponent he will face, your GPT needs what we call “fine tuning”. It is the refinement of its capabilities and functions to give you results that best meet your expectations. This is particularly crucial if you have supplied ChatGPT with your own data. This involves re- training the model on your dataset. Fine tuning allows you to adjust the model so that it performs better according to your sources or your requests. Concretely, ask him a bunch of questions and correct the answers to improve both the substance and the form of his answers.
4: An AI just for you… or to share
ChatGPT offers you the possibility of making your personalized GPT public (open to all) or sharing it only with those of your choice via a link. The service can also be integrated into a site, for example. This then requires a little more advanced knowledge (use of an API or a backend server), but it is doable. Initially, only making it available to a small group of people to receive
their opinions and constructive criticism is recommended.
It is possible to access a list of public customized GPTs to add them to your space and use them. Above all, keep in mind that the data you may have sent to GPT is not ultra-secure. We advise against the use of sensitive data (accounting, customer files, password list) in this context. Finally, be aware of potential biases (if you only train your GPT with a certain type of data, the
answers will be modeled on this specificity).
5: Provide after-sales service
Here you are with a personalized GPT, but that does not mean that the tool is fixed. Constantly monitor real-world model performance and collect user feedback. Make updates and adjustments to the model based on this feedback and developments in the application domain. ChatGPT evolves quickly, very quickly, so you might as well take full advantage of the multiple
improvements that appear over time.
