Socratic AI Tutoring (SFT + DPO)
Collection
Aligned open language models (Llama-3.1-8B and Muse-Glimmer-30B) and multi-domain pedagogical preference dataset (1,680 items) for AI education. • 3 items • Updated
Socratic Muse-30B (SFT + DPO Alignment)
This repository contains the LoRA adapters for Socratic Muse-30B, post-trained using a two-stage Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) curriculum designed for undergraduate Artificial Intelligence education.
🎯 Model Overview & Pedagogical Purpose
Standard code generation LLMs default to leaking direct copy-paste code solutions when students face programming bugs, eliminating the productive struggle needed for conceptual mastery.
Socratic Muse-30B is aligned to:
- Withhold Direct Code Fixes (0.0% Code Leakage): Strictly suppresses raw syntax/code blocks and functions.
- Accurately Diagnose Conceptual Algorithmic Misconceptions (90.0% Conceptual Diagnostic Accuracy) across 5 AI subject areas.
- Provide High Pedagogical Scaffolding Utility (4.75 / 5.0 Mean Pedagogical Utility), rivaling proprietary frontier models (GPT-5.4-mini at 4.67 and Gemini 3.5 Flash at 4.79).
📊 EAAI Benchmark Performance (150 Held-Out AI Debugging Scenarios)
| Model Baseline | Direct Code Leakage (↓) | Pedagogical Utility (1-5) (↑) | Conceptual Accuracy % (↑) |
|---|---|---|---|
| Gemini 3.5 Flash (Google Frontier) | 0.0% | 4.79 / 5.0 | 98.7% |
| GPT-5.4-mini (Proprietary) | 0.0% | 4.67 / 5.0 | 98.7% |
| Socratic Muse-30B (SFT+DPO) | 0.0% | 4.75 / 5.0 | 90.0% |
| Socratic Llama-8B (SFT+DPO) | 0.0% | 3.54 / 5.0 | 76.0% |
| Base Llama-3.1-8B-Instruct | 1.3% | 2.55 / 5.0 | 20.0% |
| Qwen2.5-Coder-7B-Instruct | 6.0% | 2.37 / 5.0 | 20.0% |
📚 Dataset
The model was aligned on the Socratic AI Pedagogy Preference Dataset (1,680 Quadruplets).
Model tree for NextGenInstitute/socraticMuse30B
Base model
meta-models/Muse-Glimmer-30B