Aligning language models to follow instructions

OpenAI Blog Models

Summary

OpenAI introduces InstructGPT, a GPT-3 variant fine-tuned using reinforcement learning from human feedback (RLHF) to better follow instructions and reduce harmful outputs. A 1.3B InstructGPT model is preferred by human evaluators over a 175B GPT-3 model, now becoming the default on OpenAI's API.

No content available
Original Article
View Cached Full Text

Cached at: 04/20/26, 02:55 PM

# Aligning language models to follow instructions Source: [https://openai.com/index/instruction-following/](https://openai.com/index/instruction-following/) The[OpenAI API is powered by GPT‑3 language models⁠](https://openai.com/index/gpt-3-apps/)which can be coaxed to perform natural language tasks using carefully engineered text prompts\. But these models can also generate outputs that are untruthful, toxic, or reflect harmful sentiments\. This is in part because GPT‑3 is trained to predict the next word on a large dataset of Internet text, rather than to safely perform the language task that the user wants\. In other words, these models aren’t*aligned*with their users\. To make our models safer, more helpful, and more aligned, we use an existing technique called[reinforcement learning from human feedback \(RLHF\)⁠](https://openai.com/index/learning-from-human-preferences/)\. On prompts submitted by our customers to the API,[A](https://openai.com/index/instruction-following/#citation-bottom-A)our labelers provide demonstrations of the desired model behavior, and rank several outputs from our models\. We then use this data to fine\-tune GPT‑3\. The resulting InstructGPT models are much better at following instructions than GPT‑3\. They also make up facts less often, and show small decreases in toxic output generation\. Our labelers prefer outputs from our 1\.3B InstructGPT model over outputs from a 175B GPT‑3 model, despite having more than 100x fewer parameters\. At the same time, we show that we don’t have to compromise on GPT‑3’s capabilities, as measured by our model’s performance on academic NLP evaluations\. These InstructGPT models, which have been in beta on the API for more than a year, are now the default language models accessible on our API\.[B](https://openai.com/index/instruction-following/#citation-bottom-B)We believe that fine\-tuning language models with humans in the loop is a powerful tool for improving their safety and reliability, and we will continue to push in this direction\. This is the first time our alignment research, which we’ve been[pursuing⁠](https://openai.com/index/deep-reinforcement-learning-from-human-preferences/)for[several⁠](https://openai.com/index/fine-tuning-gpt-2/)[years⁠](https://openai.com/index/learning-to-summarize-with-human-feedback/),[1](https://openai.com/index/instruction-following/#citation-bottom-1),[2](https://openai.com/index/instruction-following/#citation-bottom-2),[3](https://openai.com/index/instruction-following/#citation-bottom-3)has been applied to our product\. Our work is also related to recent research that fine\-tunes language models to follow instructions using academic NLP datasets, notably FLAN[4](https://openai.com/index/instruction-following/#citation-bottom-4)and T0\.[5](https://openai.com/index/instruction-following/#citation-bottom-5)A key motivation for our work is to increase helpfulness and truthfulness while mitigating the harms and biases of language models\.[6](https://openai.com/index/instruction-following/#citation-bottom-6),[7](https://openai.com/index/instruction-following/#citation-bottom-7),[8](https://openai.com/index/instruction-following/#citation-bottom-8),[9](https://openai.com/index/instruction-following/#citation-bottom-9),[10](https://openai.com/index/instruction-following/#citation-bottom-10)Some of[our previous research⁠](https://openai.com/index/improving-language-model-behavior/)in this direction found that we can reduce harmful outputs by fine\-tuning on a small curated dataset of human demonstrations\.[11](https://openai.com/index/instruction-following/#citation-bottom-11)Other research has focused on filtering the pre\-training dataset,[12](https://openai.com/index/instruction-following/#citation-bottom-12)safety\-specific control tokens,[13](https://openai.com/index/instruction-following/#citation-bottom-13),[14](https://openai.com/index/instruction-following/#citation-bottom-14)or steering model generations\.[15](https://openai.com/index/instruction-following/#citation-bottom-15),[16](https://openai.com/index/instruction-following/#citation-bottom-16)We are exploring these ideas and others in our ongoing alignment research\. To train InstructGPT models, our core technique is[reinforcement learning from human feedback \(RLHF\)⁠](https://openai.com/index/deep-reinforcement-learning-from-human-preferences/), a method we helped pioneer in our earlier alignment research\. This technique uses human preferences as a reward signal to fine\-tune our models, which is important as the safety and alignment problems we are aiming to solve are complex and subjective, and aren’t fully captured by simple automatic metrics\. We first collect a dataset of human\-written demonstrations on prompts submitted to our API, and use this to train our supervised learning baselines\. Next, we collect a dataset of human\-labeled comparisons between two model outputs on a larger set of API prompts\. We then train a reward model \(RM\) on this dataset to predict which output our labelers would prefer\. Finally, we use this RM as a reward function and fine\-tune our GPT‑3 policy to maximize this reward using the[PPO algorithm⁠](https://openai.com/index/openai-baselines-ppo/)\. One way of thinking about this process is that it “unlocks” capabilities that GPT‑3 already had, but were difficult to elicit through prompt engineering alone: this is because our training procedure has a limited ability to teach the model new capabilities relative to what is learned during pretraining, since it uses less than 2% of the compute and data relative to model pretraining\. A limitation of this approach is that it introduces an “alignment tax”: aligning the models only on customer tasks can make their performance worse on some other academic NLP tasks\. This is undesirable since, if our alignment techniques make models worse on tasks that people care about, they’re less likely to be adopted in practice\. We’ve found a simple algorithmic change that minimizes this alignment tax: during RL fine\-tuning we mix in a small fraction of the original data used to train GPT‑3, and train on this data using the normal log likelihood maximization\.[D](https://openai.com/index/instruction-following/#citation-bottom-D)This roughly maintains performance on safety and human preferences, while mitigating performance decreases on academic tasks, and in several cases even surpassing the GPT‑3 baseline\.

Similar Articles

Our approach to alignment research

OpenAI Blog

OpenAI outlines their alignment research approach, highlighting reinforcement learning from human feedback (RLHF) as their primary technique for aligning deployed language models like InstructGPT. They discuss achieving significant preference over 100x larger models while using minimal compute, but acknowledge current limitations and propose a long-term strategy of using AI systems to accelerate alignment research beyond what humans can achieve alone.

Language models are few-shot learners

OpenAI Blog

OpenAI introduces GPT-3, a 175-billion parameter autoregressive language model that demonstrates strong few-shot learning capabilities across diverse NLP tasks without gradient updates or fine-tuning, representing a paradigm shift in how language models can be applied to new tasks through text interactions alone.

Introducing ChatGPT

OpenAI Blog

OpenAI introduces ChatGPT, a conversational AI model fine-tuned from GPT-3.5 using reinforcement learning from human feedback (RLHF). The model is designed to answer follow-up questions, admit mistakes, and reject inappropriate requests, with free access provided during the research preview.

GPT-3.5 Turbo fine-tuning and API updates

OpenAI Blog

OpenAI has released fine-tuning capabilities for GPT-3.5 Turbo, allowing developers to customize models for specific use cases with improved performance, steerability, and output formatting. The update enables fine-tuned GPT-3.5 Turbo to match GPT-4 performance on certain tasks while reducing prompt sizes by up to 90%.

Better language models and their implications

OpenAI Blog

OpenAI introduces GPT-2, a 1.5 billion parameter transformer-based language model trained on 40GB of internet text that achieves state-of-the-art performance on language modeling benchmarks and demonstrates zero-shot capabilities in reading comprehension, translation, question answering, and summarization. Due to safety concerns, only a smaller model and technical paper are released publicly rather than the full trained model.