Model Card for ML-GT/Llama-3.3-70B-Instruct-Edubuddy
EduBuddy is a Socratic AI Teaching Assistant designed for Georgia Tech’s CS 4641/7641 Machine Learning course.
It integrates fine-tuning and retrieval-augmented generation (RAG) to replicate TA-style pedagogical communication — guiding students through reasoning rather than providing direct answers.
Model Details
Model Description
EduBuddy fine-tunes Llama-3.3-70B-Instruct to emulate the Socratic teaching style observed in 8,197 authentic student–TA conversations from five years of Georgia Tech ML coursework.
It uses a dual-RAG architecture to retrieve:
- Course-specific lecture and homework materials.
- Structured learning subgoals for scaffolding conceptual understanding.
EduBuddy’s responses are contextually grounded, Socratic in nature, and aligned with the structure of the ML course’s assignments and lectures.
- Developed by: Georgia Tech ML TA Team
- Model type: Instruction-tuned large language model (LLM) fine-tuned for pedagogical dialogue
- Language(s): English
- License: No license (research-only use)
- Finetuned from:
meta-llama/Llama-3.3-70B-Instruct
- Intended users: Instructors, researchers, and educational technologists exploring Socratic AI tutoring
- Primary domain: University-level Machine Learning education (Georgia Tech CS 4641/7641)
Model Sources
- Paper: EduBuddy: A Socratic AI Teaching Assistant combining Fine-Tuning and Retrieval-Augmented Generation for ML Education
- Repository: [Not released publicly]
- Demo: None (research prototype only)
Uses
Direct Use
- To assist ML students by asking guiding questions and offering step-by-step reasoning in conceptual and coding queries.
- To emulate TA responses for course-related Q&A forums or tutoring systems.
Downstream Use
- Research on Socratic dialogue systems, AI pedagogy, or educational LLM design.
- Fine-tuning base models for similar instructional contexts.
Out-of-Scope Use
- Providing direct answers or code solutions.
- Use outside educational contexts (e.g., general tutoring, grading automation, or production deployment).
Bias, Risks, and Limitations
- Course specificity: Designed for Georgia Tech ML course materials; may perform poorly on unrelated topics.
- Data artifacts: Training data includes informal language, emojis, and stylistic variation from real forums.
- Scale dependency: Smaller models (e.g., 8B) underperform significantly; pedagogical reasoning emerges reliably only at large scale (70B+).
- RAG noise sensitivity: Retrieval documents can introduce irrelevant context when not aligned with queries.
Recommendations
Users should treat EduBuddy as a pedagogical research prototype, not a general-purpose tutor.
Before adaptation to new domains, retraining on course-specific dialogues and materials is recommended.
