Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Open Source Language Models Can Provide Feedback: Evaluating LLMs' Ability to Help Students Using GPT-4-As-A-Judge

Charles Koutcheme, Nicola Dainese, Sami Sarsa, Arto Hellas, Juho Leinonen, Paul Denny

TL;DR

This work investigates whether open-source language models can generate useful feedback for student programming work and whether GPT-4 can serve as a reliable judge of that feedback. Using a Dart course dataset and a two-stage evaluation (GPT-4 as judge of GPT-3.5 feedback, then open-source models generating feedback evaluated by GPT-4), the study reveals a moderate alignment with human judgments and a tendency for GPT-4 to be optimistic. Zephyr-7B open-source models demonstrate competitive feedback quality relative to GPT-3.5, narrowing the gap to GPT-4 in several cases, and highlighting privacy and cost advantages for education. Overall, the results support the potential of open-source feedback generation in classrooms while emphasizing that human evaluation remains important for per-feedback quality assurance. The work also contributes toward more accessible, privacy-preserving educational tools by releasing code and outlining directions for future improvements.

Abstract

Large language models (LLMs) have shown great potential for the automatic generation of feedback in a wide range of computing contexts. However, concerns have been voiced around the privacy and ethical implications of sending student work to proprietary models. This has sparked considerable interest in the use of open source LLMs in education, but the quality of the feedback that such open models can produce remains understudied. This is a concern as providing flawed or misleading generated feedback could be detrimental to student learning. Inspired by recent work that has utilised very powerful LLMs, such as GPT-4, to evaluate the outputs produced by less powerful models, we conduct an automated analysis of the quality of the feedback produced by several open source models using a dataset from an introductory programming course. First, we investigate the viability of employing GPT-4 as an automated evaluator by comparing its evaluations with those of a human expert. We observe that GPT-4 demonstrates a bias toward positively rating feedback while exhibiting moderate agreement with human raters, showcasing its potential as a feedback evaluator. Second, we explore the quality of feedback generated by several leading open-source LLMs by using GPT-4 to evaluate the feedback. We find that some models offer competitive performance with popular proprietary LLMs, such as ChatGPT, indicating opportunities for their responsible use in educational settings.

Open Source Language Models Can Provide Feedback: Evaluating LLMs' Ability to Help Students Using GPT-4-As-A-Judge

TL;DR

This work investigates whether open-source language models can generate useful feedback for student programming work and whether GPT-4 can serve as a reliable judge of that feedback. Using a Dart course dataset and a two-stage evaluation (GPT-4 as judge of GPT-3.5 feedback, then open-source models generating feedback evaluated by GPT-4), the study reveals a moderate alignment with human judgments and a tendency for GPT-4 to be optimistic. Zephyr-7B open-source models demonstrate competitive feedback quality relative to GPT-3.5, narrowing the gap to GPT-4 in several cases, and highlighting privacy and cost advantages for education. Overall, the results support the potential of open-source feedback generation in classrooms while emphasizing that human evaluation remains important for per-feedback quality assurance. The work also contributes toward more accessible, privacy-preserving educational tools by releasing code and outlining directions for future improvements.

Abstract

Large language models (LLMs) have shown great potential for the automatic generation of feedback in a wide range of computing contexts. However, concerns have been voiced around the privacy and ethical implications of sending student work to proprietary models. This has sparked considerable interest in the use of open source LLMs in education, but the quality of the feedback that such open models can produce remains understudied. This is a concern as providing flawed or misleading generated feedback could be detrimental to student learning. Inspired by recent work that has utilised very powerful LLMs, such as GPT-4, to evaluate the outputs produced by less powerful models, we conduct an automated analysis of the quality of the feedback produced by several open source models using a dataset from an introductory programming course. First, we investigate the viability of employing GPT-4 as an automated evaluator by comparing its evaluations with those of a human expert. We observe that GPT-4 demonstrates a bias toward positively rating feedback while exhibiting moderate agreement with human raters, showcasing its potential as a feedback evaluator. Second, we explore the quality of feedback generated by several leading open-source LLMs by using GPT-4 to evaluate the feedback. We find that some models offer competitive performance with popular proprietary LLMs, such as ChatGPT, indicating opportunities for their responsible use in educational settings.
Paper Structure (19 sections, 4 figures, 1 table)

This paper contains 19 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. Automatic Feedback in Computing Education
  4. Using Language Models as Judges
  5. Methods
  6. Data
  7. Generating Feedback
  8. Prompting.
  9. Models.
  10. Automatic Feedback Evaluation
  11. Results and Discussion
  12. Automatic Feedback Evaluation
  13. Sources of misjudgment.
  14. Research and practical implications.
  15. Open-source judges.
  16. ...and 4 more sections

Figures (4)

  • Figure 1: Feedback prompt template. We provide (1) a system prompt specifying the behaviour of the model, (2) a description of the feedback task, and (3) contextual information.
  • Figure 2: Judging prompt template. We provide (1) a system prompt specifying GPT-4's behaviour, (2) a description of the grading task, and (3) contextual information.
  • Figure 3: Fraction of 'comprehensive' (i.e., satisfying all three criteria) and 'insightful' (i.e., perceptive and selective) feedback for all language models.
  • Figure 4: Examples of generated feedback. We present a problem description together with the incorrect code solution, the feedback generated by a proprietary model (GPT-3.5) and the feedback generated by a open-source model (Zephyr-7B). Problem descriptions, and strings and variable names in the examples shown have been translated from the original language (Finnish); LLM feedback was in English and thus not translated.