EnTraineR

EnTraineR logo

An intelligent teaching assistant based on LLMs to help interpret statistical model outputs in R.
EnTraineR builds audience-aware prompts (beginner, applied, advanced) that never invent numbers: it passes verbatim outputs from R and instructs how to explain them.

Installation

From GitHub:

# install.packages("remotes")
# remotes::install_github("Sebastien-Le/EnTraineR")
install.packages("EnTraineR", repos = c('https://sebastien-le.r-universe.dev', 'https://cloud.r-project.org'))

Optional but recommended packages for examples:

What it does

Included datasets

The package ships 3 small datasets for teaching:

These datasets are the intellectual property of L'Institut Agro Rennes Angers and are used for the "Statistical Approach" course module.

data(deforestation); str(deforestation)
data(ham); summary(ham)
data(poussin); with(poussin, table(Temperature, Gender))

Quick start

1) ANOVA (AovSum)

# install.packages("SensoMineR")
library(SensoMineR)
data(chocolates)

# Build AovSum (example similar to chocolates::Granular ~ Product*Panelist)
res <- AovSum(Granular ~ Product*Panelist, data = sensochoc)

intro <- "Six chocolates have been evaluated by a sensory panel, 
  according to a sensory attribute: granular.
  The panel has been trained according to this attribute
  and panellists should be reproducible when rating this attribute."
intro <- gsub("\n", " ", intro)
intro <- stringr::str_squish(intro)

p <- trainer_aovsum(
  aovsum_obj   = res,
  audience     = "applied",
  t_test       = c("Product", "Panelist"),  # filter T-test section
  introduction = intro
)

cat(p)   # a ready-to-use prompt for an LLM or for teaching

2) Linear model (FactoMineR::LinearModel)

# install.packages("FactoMineR"); install.packages("stringr")
library(FactoMineR)

intro_ham <- "Can we predict ham overall liking from its sensory profile?"
intro_ham <- stringr::str_squish(gsub("\n", " ", intro_ham))

fit <- LinearModel(`Overall liking` ~ ., data = ham, selection = "bic")

pr <- trainer_linear_model(
  lm_obj       = fit,
  introduction = intro_ham,
  audience     = "advanced"
)

cat(pr)

Another linear model with interaction and a categorical factor:

fit2 <- LinearModel(Temp_water ~ Temp_air * Deforestation,
                    data = deforestation, selection = "none")

pr2 <- trainer_linear_model(
  lm_obj       = fit2,
  introduction = "Effect of deforestation on the air-water temperature link.",
  audience     = "beginner"
)

cat(pr2)

3) Classical tests

t-test:

tt <- t.test(rnorm(20, 0.1), mu = 0)
cat(trainer_t_test(tt, audience = "beginner"))

Variance F-test:

vt <- var.test(rnorm(25, sd = 1.0), rnorm(30, sd = 1.3))
cat(trainer_var_test(vt, audience = "applied"))

Proportion test:

pt <- prop.test(x = c(42, 35), n = c(100, 90))
cat(trainer_prop_test(pt, audience = "advanced", summary_only = TRUE))

Correlation test:

set.seed(1)
x <- rnorm(30); y <- 0.5 * x + rnorm(30, sd = 0.8)
ct <- cor.test(x, y, method = "pearson")
cat(trainer_cor_test(ct, audience = "applied"))

Chi-squared test:

m <- matrix(c(10, 20, 30, 40), nrow = 2)
cx <- chisq.test(m, correct = TRUE)
cat(trainer_chisq_test(cx, audience = "beginner"))

Using Gemini from R (optional)

gemini_generate() lets you send a prompt to Google Gemini and get the response back as text.

# 1) Set your API key once per session (or in .Renviron)
Sys.setenv(GEMINI_API_KEY = "your_key_here")

# 2) Send a prompt
txt <- gemini_generate(
  prompt      = "Say hello in one short sentence.",
  model       = "gemini-2.5-flash",   # accepts "gemini-2.5-flash" or "models/gemini-2.5-flash"
  temperature = 0.2,
  user_agent  = "EnTraineR/0.9.0 (https://github.com/Sebastien-Le/EnTraineR)"
)
cat(txt)

Audience profiles (summary)

All prompts emphasize: do not invent numbers; use only what appears in the printed output.

Reproducibility and LLMs

By default, trainers return a prompt string (i.e., generate = FALSE).
If you have a generator backend, you can pass generate = TRUE and a llm_model name; implement your own trainer_core_generate_or_return() to call your LLM API.

Contributing

Issues and pull requests are welcome. Please:

License and citation

See the DESCRIPTION file for license terms.
If EnTraineR helps your teaching or analyses, starring the repo is appreciated.

Acknowledgments

Thanks to the R community and the authors of FactoMineR and SensoMineR for inspiring teaching tools and example datasets used in demonstrations.