June 2018
OpenAI released the GPT-1 model with 110 million parameters.
In November 2018
OpenAI released the GPT-2 model with 1.5 billion parameters, but due to concerns about abuse, all the code and data of the model are not open to the public.
February 2019
OpenAI opened some code and data of the GPT-2 model, but access is still restricted.
June 10, 2019
OpenAI released the GPT-3 model with 175 billion parameters and provided access to some partners.
September 2019
OpenAI opened all the code and data of GPT-2 and released a larger version.
In May 2020
OpenAI announced the launch of the beta version of the GPT-3 model, which has 175 billion parameters and is the largest natural language processing model to date.
March 2022
OpenAI released InstructGPT, using Instruction Tuning
November 30, 2022
OpenAI passed the GPT-3.5 series of large-scale languages The new conversational AI model ChatGPT is officially released after fine-tuning the model.
December 15, 2022
The first update to ChatGPT, which improves overall performance and adds new features to save and view historical conversation records.
January 9, 2023
ChatGPT was updated for the second time, improving the authenticity of answers and adding a new "stop generation" function.
January 21, 2023
OpenAI released a paid version of ChatGPT Professional that is limited to some users.
January 30, 2023
The third update of ChatGPT not only improves the authenticity of answers, but also improves mathematical skills.
February 2, 2023
OpenAI officially launched the ChatGPT paid version subscription service. The new version responds faster and runs more stably than the free version.
March 15, 2023
OpenAI shockingly launched the large-scale multi-modal model GPT-4, which can not only read text, but also recognize images and generate text results. It is now connected ChatGPT is open to Plus users.
Before the emergence of GPT, NLP models were mainly trained based on large amounts of annotated data for specific tasks. This will lead to some limitations:
Large-scale high-quality annotation data is not easy to obtain;
The model is limited to the training it has received, and its generalization ability is insufficient;
It cannot be executed Out-of-the-box tasks limit the practical application of the model.
In order to overcome these problems, OpenAI embarked on the path of pre-training large models. GPT-1 is the first pre-trained model released by OpenAI in 2018. It adopts a one-way Transformer model and uses more than 40GB of text data for training. The key features of GPT-1 are: generative pre-training (unsupervised) and discriminative task fine-tuning (supervised). First, we used unsupervised learning pre-training and spent 1 month on 8 GPUs to enhance the language capabilities of the AI system from a large amount of unlabeled data and obtain a large amount of knowledge. Then we conducted supervised fine-tuning and compared it with large data sets. Integrated to improve system performance in NLP tasks. GPT-1 showed excellent performance in text generation and understanding tasks, becoming one of the most advanced natural language processing models at the time.
Since the single-task model lacks generalization and multi-task learning requires a large number of effective training pairs, GPT-2 is It has been expanded and optimized on the basis of GPT-1, removing supervised learning and only retaining unsupervised learning. GPT-2 uses larger text data and more powerful computing resources for training, and the parameter size reaches 150 million, far exceeding the 110 million parameters of GPT-1. In addition to using larger data sets and larger models to learn, GPT-2 also proposes a new and more difficult task: zero-shot learning (zero-shot), which is to directly apply pre-trained models to many downstream Task. GPT-2 has demonstrated excellent performance on multiple natural language processing tasks, including text generation, text classification, language understanding, etc.
GPT-3 is the latest in the GPT series of models A model that uses a larger parameter scale and richer training data. The parameter scale of GPT-3 reaches 1.75 trillion, which is more than 100 times that of GPT-2. GPT-3 has shown amazing capabilities in natural language generation, dialogue generation and other language processing tasks. In some tasks, it can even create new forms of language expression.
GPT-3 proposes a very important concept: In-context learning. The specific content will be explained in the next tweet.
The training of InstructGPT/ChatGPT is divided into 3 steps, and the data required for each step is slightly different. We will introduce them separately below.
Start with a pre-trained language model and apply the following three steps.
Step 1: Supervised fine-tuning SFT: Collect demonstration data and train a supervised policy. Our tagger provides a demonstration of the desired behavior on the input prompt distribution. We then use supervised learning to fine-tune the pre-trained GPT-3 model on these data.
Step 2: Reward Model training. Collect comparative data and train a reward model. We collected a dataset of comparisons between model outputs, where labelers indicate which output they prefer for a given input. We then train a reward model to predict human-preferred outputs.
Step 3: Reinforcement learning via proximal policy optimization (PPO) on the reward model: use the output of the RM as the scalar reward. We use the PPO algorithm to fine-tune the supervision strategy to optimize this reward.
Steps 2 and 3 can be iterated continuously; more comparison data is collected on the current optimal strategy, which is used to train a new RM, and then a new strategy.
The prompts in the first two steps come from user usage data on OpenAI’s online API and handwritten by hired annotators. The last step is all sampled from the API data. The specific data of InstructGPT:
1. SFT data set
The SFT data set is used to train the first step The supervised model uses the new data collected to fine-tune GPT-3 according to the training method of GPT-3. Because GPT-3 is a generative model based on prompt learning, the SFT dataset is also a sample composed of prompt-reply pairs. Part of the SFT data comes from users of OpenAI’s PlayGround, and the other part comes from the 40 labelers employed by OpenAI. And they trained the labeler. In this dataset, the annotator's job is to write instructions themselves based on the content.
2. RM data set
The RM data set is used to train the reward model in step 2. We also need to set a reward target for the training of InstructGPT/ChatGPT. This reward goal does not have to be differentiable, but it must align as comprehensively and realistically as possible with what we need the model to generate. Naturally, we can provide this reward through manual annotation. Through artificial pairing, we can give lower scores to the generated content involving bias to encourage the model not to generate content that humans do not like. The approach of InstructGPT/ChatGPT is to first let the model generate a batch of candidate texts, and then use the labeler to sort the generated content according to the quality of the generated data.
3. PPO data set
The PPO data of InstructGPT is not annotated. It comes from GPT-3 API users. There are different types of generation tasks provided by different users, with the highest proportion including generation tasks (45.6%), QA (12.4%), brainstorming (11.2%), dialogue (8.4%), etc.
Appendix:
Sources of various capabilities of ChatGPT:
The above is the detailed content of ChatGPT topic one GPT family evolution history. For more information, please follow other related articles on the PHP Chinese website!
![[Web front-end] Node.js quick start](https://img.php.cn/upload/course/000/000/067/662b5d34ba7c0227.png)
