ChatGPT, which stands for Chat Generative Pre-trained Transformer, is a chatbot developed by OpenAI. ChatGPT is built on top of OpenAI's GPT-3.5 family of large language models, and is fine-tuned with both supervised and reinforcement learning techniques.
ChatGPT was launched as a prototype in November 2022, and quickly garnered attention for its detailed responses and articulate answers across many domains of knowledge.
It was fine-tuned on top of GPT-3.5 using supervised learning as well as reinforcement learning. Both approaches used human trainers to improve the model's performance. In the case of supervised learning, the model was provided with conversations in which the trainers played both sides: the user and the AI assistant. In the reinforcement step, human trainers first ranked responses that the model had created in a previous conversation. These rankings were used to create 'reward models' that the model was further fine-tuned on using several iterations of Proximal Policy Optimization (PPO). Proximal Policy Optimization algorithms present a cost-effective benefit to trust region policy optimization algorithms; they negate many of the computationally expensive operations with faster performance. The models were trained in collaboration with Microsoft on their Azure supercomputing infrastructure.
ChatGPT is part of a series of releases around GPT 3.5 that are highlighting some of the capabilities of the upcoming GPT-4 model. One of the key differences of ChatGPT with previous models is its ability to follow instructions. This is powered another model called InstructGPT which OpenAI quietly unveiled at the beginning of the year.
Large language models like GPT-3 are often used to follow instructions to execute user’s tasks. However, quite often, these models generate toxic or untruthful outputs that are not related to the input instructions. This is mostly due to the fact that models like GPT-3 are trained to predict the next word in a sentence rather than to execute a specific task. This is precisesly the problem OpenAI tried to address with InstructGPT, a language model that builds upon GPT-3 language capabilities but improves it its ability to follow instructions.
More about InstructGPT
Training language models to follow instructions with human feedback.
Making language models bigger does not inherently make them better at following a user's intent. For example, large language models can generate outputs that are untruthful, toxic, or simply not helpful to the user. In other words, these models are not aligned with their users. In this paper, we show an avenue for aligning language models with user intent on a wide range of tasks by fine-tuning with human feedback. Starting with a set of labeler-written prompts and prompts submitted through the OpenAI API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to fine-tune GPT-3 using supervised learning. We then collect a dataset of rankings of model outputs, which we use to further fine-tune this supervised model using reinforcement learning from human feedback. We call the resulting models InstructGPT. In human evaluations on our prompt distribution, outputs from the 1.3B parameter InstructGPT model are preferred to outputs from the 175B GPT-3, despite having 100x fewer parameters. Moreover, InstructGPT models show improvements in truthfulness and reductions in toxic output generation while having minimal performance regressions on public NLP datasets. Even though InstructGPT still makes simple mistakes, our results show that fine-tuning with human feedback is a promising direction for aligning language models with human intent.
Post
