Large language models are not a panacea, and due to the lack of industry expertise, large models actually have certain limitations in solving practical problems. So, how to make the big model understand you better? As the designer of the AI development platform, the author summarizes two simple and efficient methods, let's take a look at them together.
We all know that the emergence of generative AI has set off a wave of artificial intelligence, and in this context, understanding the development methods behind AI products can help us better use AI products.
1. What is generative AI?
Generative AI can help us do a lot of things, and in our daily lives, it can be used to generate reports and improve the efficiency of reporting; In the field of e-commerce, intelligent customer service can automatically answer and solve users' questions; In the medical field, intelligent doctors can help patients diagnose diseases, improving the work efficiency of industry personnel.
2. The essence of ChatGPT is the application of large-scale language technology
Among the generative AI applications, we are familiar with Open AI's ChatGPT, Baidu's Wenxin Yiyan, Byte's Doubao, etc., their essence is to apply a large model technology.
This technology is obtained by professional technicians in cloud vendors through massive text data processing and expensive computing costs. Such a technique allows large models to learn human language patterns and knowledge structures and generate natural and fluent responses.
The large model is analogous to the brain
Popular analogy: We can analogy large models to "brains", a large amount of text data is equivalent to "information provided by the outside world", and expensive algorithms are likened to "hired senior professors", which convert external information into knowledge points, store them in the brain, and finally present the smart products we use.
Third, large language models are not a panacea, and there are limitations in solving practical problems
However, large language models are not a panacea, and they have certain limitations in solving practical problems due to a lack of industry domain expertise.
For example, in the following scenario: I want to understand the relevant design rules of CXD intelligent cloud products through Wenxin Yiyan, but its answer cannot solve my actual problem.
Wenxin Yiyan only provides design rules that apply to general platforms, and these rules are not fully applicable to our products. What should I do if I encounter this problem in the process of enterprise application?
Fourth, the solution: let the artificial intelligence application understand you better and get the content you want
As a designer of an AI development platform, I have summarized two simple and efficient methods to share with you.
Method 1 "Teach it to find"
The principle of "teaching it to find" is: by adding prompt words, using the key words in it, prompting the large model to understand our intention, and finding the information we want in the existing data information and then answering.
This method can help us improve the effect of large models at a low cost and quickly.
Take the development of a "car sales customer service" as a scenario, and the Diffy product as a tool demonstration.
The left panel of the product is the configuration of relevant parameters, and the right side is the test tool for the user's real use scenarios.
Without a prompt configuration, I chose ChatGPT 3.5 to answer my question, and found that its answer did not have substantive and valid information, which was not very helpful to the user who bought the car.
How can I improve quality by adding prompts?
Step 1: Add prompt words: Add text to the input box on the left side of the white to restrict the large model's answers, so that the large model can provide comparison information for the performance and appearance of the car as a professional salesperson.
Step 2: Test the effect: ChatGPT's answer effect is significantly improved.
In this way, it can be packaged into a new application for users to use, and users can ask questions and answers within a large model with a defined range, which can greatly improve product satisfaction.
Based on this method, I recommend two types of easy-to-use tools for you: one is the prompt template platform, in which you can get high-quality prompts from all walks of life, education, finance, etc., which can be copied and used directly.
Dify：http://cloud.dify.ai/explore/apps
GPT short：http://prompt-shortcut.writeathon.cn
Thousand Sails Large Model Platform: http://prompt-shortcut.writeathon.cn
Method 2 "Teach Him to Learn"
The principle of "teaching others to learn" is: by adding its own data, combining it with its own general data, teaching it to learn new knowledge, so as to customize a new large model that understands you.
How do I add my own data? There are two ways.
Method 1: You can generate a new large model from unstructured documents (such as pdf word documents, web links, etc.) on an AI customized platform.
Method 2: Select an AI development platform, prepare a structured dataset (text-to-information) excel json file, and retrain the large model to learn new knowledge.
Again, I'll use two examples to illustrate how this works.
Case 1: Using the development of the "Design Specification Assistant" as a scenario, the Chatbase platform is demonstrated
Step 1: Select Create bot.
Step 2: Select data, there are 5 types of data provided, which are documents, text, connections, Q&A pairs, and third-party note URLs.
I have prepared three kinds of documents for daily use in advance, namely: design specification documents, scheme library documents, and research reports for each product, and at the same time added professional prompt texts and links to the official website of the cloud design center platform to help the large model better learn our knowledge.
Step 3: Build the bot and start testing.
Through three rounds of conversations, it easily answered the questions I wanted, provided the correct normative information and health indicators, and summarized information on how to design a data labeling scenario.
Finally, all you need to do is publish it as a website and provide the link to the designers in the group.
Case 2: The development of the "medical customer service assistant" scenario is demonstrated on the Qianfan platform.
Step 1: Prepare the data, you can choose to prepare the dataset by yourself, and follow the example guidance of the platform to operate; You can also directly use the ready-made industry datasets provided by the platform. For example, we chose this medical Chinese dataset.
Step 2: Train the model, in this page, first select the trained large model, which can be selected according to the introduction of the large model. We chose Ernie bot because it works better in Chinese.
The second is the selection of training methods, which affect the amount of resource cost, consumption time and model stability. Finally, the training parameters are selected, and the platform will provide recommended values based on the information that has been provided, and if the developer has an understanding of the training parameters, they can adjust them based on experience. Click OK to start training,
Step 3: After the training is completed, publish the task as a model and deploy the model as a service. (I won't expand on it in detail here)
Step 4: In the experience center, select the service you just deployed, and you can test it to understand the actual effect of this medical customer service.
The above is the operation process of adding structured data to the large model, which requires more money and time, but is more suitable for enterprises that need high-precision answers.
1) Data service tools
These platforms provide ready-to-use datasets, so you don't have to spend time preparing data, such as wine knowledge, animal knowledge, etc., to download and use according to your business needs.
阿里modelscope：http://www.modelscope.cn/datasets? Tags=text-classification&dataType=text&page=1
百度AI Studio：http://aistudio.baidu.com/datasetoverview
DataCastle：http://www.datacastle.cn/dataset_l
2) Large model tuning tools
Foreign large-scale model tuning tools: Google, Microsoft.
Final summary
In the past, the development of large models was more done by professional technicians, but with the development of the times, AI may be beneficial to everyone in the future. In the future, we may all develop our own big model and use robots with our data to work and chat.
As AI development product designers, we have a long way to go, and we have been working the road to make the large model development process simple and easy to use~~
This article was originally published by @MINGZI on Everyone is a Product Manager and is not allowed to be reproduced without permission.
The title image is from Unsplash and is licensed under CC0.
The views in this article only represent the author's own, everyone is a product manager, and the platform only provides information storage space services.