title: "rgpt3: Making requests from R to the GPT-3 API" tags:
The past decade has seen leap advancements in the field of Natural Language Processing (NLP, i.e., using computational methods to study human language). Of particular importance are generative language models, which - among standard NLP tasks such as text classification - are able to produce text data that are often indistinguishable from human-written text. The most prominent language model is GPT-3 (short for Generative Pre-trained Transformer 3) developed by Open AI and released to the public in 2021 [@brown2020language]. While these models offer exciting potential for the study of human language at scale, models such as GPT-3 were also met with controversy [@bender2021dangers]. Part of the criticism stems from the opaque nature of the model and the potential biases it may hence propagate in generated text data. As a consequence, there is a need to understand the model and its limitations so researchers can use it in a responsible and informed manner. This package makes it possible to use the GPT-3 model from the R programming language, thereby opening access to this tool to the R community and enabling more researchers to use, test and study the powerful GPT-3 model.
The GPT-3 model has pushed the boundaries of the language abilities of artificially intelligent systems. Many tasks that were deemed unrealistic or too difficult for computational models are now deemed solvable [@vandermaas2021]. Especially the performances of the model on tasks originating from Psychology show the enormous potential of the GPT-3 model. For example, when asked to formulate creative use cases of everyday objects (e.g., a fork), the GPT-3 model produced alternative uses of the objects that were rated of higher utility (but lower originality and surprise) than creative use cases produced by human participants [@stevenson2022putting]. Others found that the GPT-3 model shows verbal behaviour similar to humans on cognitive tasks, so much so that the model made the same intuitive mistakes that are observed in humans [@binz2022using]. Aside from these efforts to understand how the model thinks, others started to understand the personality that may be represented by the model. Asked to fill in a standard personality questionnaire and a human values survey, the GPT-3 model showed a response pattern comparable with human samples and showed evidence of favouring specific values over others (e.g., self-direction > conformity) [@miotto_who_2022].
There is also ample evidence that the GPT-3 model produces biased responses (e.g., assigning attributes of brilliance more often to men than to women) [@shihadehbrilliance]. Both the promises and challenges with the GPT-3 model require that we start to understand the system better. Of particular relevance in the ambition to study such a black box language model is the "machine behaviour" [@rahwan2019machine] approach, which harnesses research designs from psychological and social science research to map out the behaviour and processes of algorithms (e.g. GPT-3).
Since a large part of the behavioural and social science community who may be best placed to conduct such research is using the R environment, this package - as the first R access point to the GPT-3 model - could help break down barriers and increase the adoption of GPT-3 research in that community.
The rgpt3 package allows users to interact via R with the GPT-3 API to perform the two core functionalities: i) requesting text completions and ii) obtaining embeddings representations from text input.
The core idea of text completions is to provide the GPT-3 model with prompts which it uses as context to generate a sequence of arbitrary length. For example, prompts may come in the form of questions (e.g., 'How does the US election work?'), tasks (e.g., 'Write a diary entry of a professional athlete:'), or open sequences that the model should finish (e.g., 'Maria has started a job as a').
This package handles completions in the most efficient manner from a data.table or data.frame object with the gpt3_completions() function. In the example, we provide the prompts from a data.frame and ask the function to produce five completions (via the param_n parameter) with a maximum token length each of 50 (param_max_tokens) with a sampling temperature of 0.8 (param_temperature). Full detail on all available function parameters is provided in the help files (e.g., ?gpt3_completions)
The output object contains a list with two data.tables: the text generations and the meta-information about the request made.
prompt_data = data.frame(prompts = c('How does the US election work?' , 'Write a diary entry of a professional athlete: ' , 'Maria has started a job as a ') , prompt_id = 1:3) output = gpt3_completions(prompt_var = prompt_data$prompts , param_max_tokens = 50 , param_n = 5 , param_temperature = 0.8)
The second (albeit less relevant for computational social science work) functionality concerns text embeddings. An embedding representation of a document can help, for example, to calculate the similarity between two pieces of text. Embeddings can be derived as follows (using the package-provided mini travel_blog_data dataset):
data("travel_blog_data") example_data = travel_blog_data[1:5, ] embeddings = gpt3_embeddings(input_var = example_data$gpt3 , id_var = 1:nrow(example_data))
When the GPT-3 model was announced, it was controversial whether the model should be made available to the public. Open AI decided to make it available through an API that users can access with their own API key that they receive when creating an account. Using the GPT-3 API is not free (although at the time of writing, each user is provided with a small amount to get started, which is sufficient for most basic research ideas).