Summary and conclusion

Day 4, B

Michael C Sachs

Using copilot and other LLMs

Purpose:

GitHub Copilot acts as an AI pair programmer, providing real-time, context-aware code completions (“ghost text”) as you type in R scripts.

Getting started:

• https://github.com/education/students
• https://docs.posit.co/ide/user/ide/guide/tools/copilot.html

How?

• LLMs work by predicting text given a context (i.e., preceding text, instructions)
• The predictions are based on a model that was trained on a corpus, a body of text that was fed to the model during training. Copilot’s corpus is all the code on Github.
• The predictions are stochastic, and they are only as good as the corpus the model was trained on

How to use it

Generating code

• Works as a sort of autocomplete, making suggestions based on the context
• The context includes the current script and comments, but can be set to include everything in an Rstudio project
• They pop up automatically when you start typing, e.g., a function/object name

Answering questions

• Use a comment with a # q: at the beginning and a question mark at the end. Copilot’s completion should be an “answer” comment # a:.

For example, the below will be completed with an a: followed by an answer:

# q: What is the definition of standard error?

Best practices

• Communicate clearly to Copilot so it generates the right code. Use the vocabulary you learned in this class
• Review and understand suggestions before accepting them (especially important in R where statistical correctness matters). Nobody wants to help you fix LLM-generated code: it is your job to make sure that it works.
• Make sure your project context (file structure, naming conventions, data types) is clear so Copilot can leverage context.
• AI-generated code isn’t guaranteed bug- or bias-free.
• Check the current KU guidelines on reporting the use of LLMs/AI

Tips for getting good results

Think about the overall goal and break it down into smaller subtasks. Write down those subtasks using the language of programming.

• Start broad, then narrow
  • Begin with a description of the overall goal or scenario.
  • Then specify details: subtasks, inputs, outputs, constraints, edge-cases.
• Provide examples
  • Example inputs and expected outputs.
  • Example implementations or comments to give Copilot context.
• Make your intent clear
  • Mention libraries/frameworks you want to use (e.g., tidyverse, ggplot2, loops, functions).
  • State any style or performance requirements.
• Use high-quality comments
  • A well written comment above an R function like # This function takes a data frame and returns a tidy summary of each numeric column
  • Explain edge cases like missing data, factor levels, grouping, etc.
• Iterate and refine
  • Use the suggestions as a draft rather than final code.
  • Be specific about “how” not just “what”: Use dplyr’s group_by() + summarise(); handle NA values; return a tibble.
  • Consider instruction-style prompts: # Write an R function named summarise_numeric() that ... gives a specific action.

Try it out

• Try using copilot to solve one of the exercises
• Does it solve it correctly? Can you explain what it does and fix it if needed?
• Does it suggest using some functions that are new to you?
• Is it faster than doing it yourself?

What did we learn?

Project organization

• Setting up your project structure
• Documentation and readme
• Working directories and file paths

Basic programming principles

• Data structures
• Loops and conditionals
• Functions, classes, generics

Working with data

• Tidying
• Reshaping
• Merging
• Dealing with dates and strings
• Visualization

Conclusion

Where to go from here?

1. Research project management and reproducibility
   • targets package
   • project templates
   • git and/or github
2. R package development
   • Great way to share code, data, and documentation
   • Packages for personal and internal use
3. High performance computing
   • Parallelization, using computing servers
   • Using C/C++ (or other languages) in R
4. Visualization and interactivity
   • shiny package
   • Dynamic documents and graphics
   • ggplot2 custom Geoms and Stats

How to get help?

• Keep learning on your own, remember that things change
• Use the linked resources to get some background
• CRAN task views, help files, vignettes
• Google “how do I do x in R”, or Search the exact error message
• Stackoverflow, Rstudio forums, R mailing lists

Feedback

Course feedback

1. What did you like about the course?
2. What did you dislike about the course?
3. Did you knowledge about R programming improve during the course?
4. Are there topics you wanted to be covered but were not?