sugar-reflection-lora

A LoRA adapter fine-tuned on top of Mistral-7B-Instruct-v0.2 for AI-guided reflective learning in educational settings.

Built as a prototype component of the Socratic Reflection Engine — a direct MVP of the GSoC 2026 proposal "AI Reflection in the Sugar Journal" for Sugar Labs.

Purpose

This adapter fine-tunes Mistral-7B-Instruct to enforce question-only responses in reflective learning dialogues. The model is specifically trained to ask Socratic questions and never answer for the child — the critical guardrail that makes this a reflection tool rather than a homework helper.

Supported Pedagogical Frameworks

• Gibbs' Reflective Cycle (6-turn conversation)
• Kolb's Experiential Learning (4-turn)
• 5Rs Framework — Bain et al. (5-turn)

Training Details

Parameter	Value
Base model	mistralai/Mistral-7B-Instruct-v0.2
Method	LoRA via HuggingFace PEFT
Quantization	4-bit (bitsandbytes)
Training steps	60
Epochs	3
Final training loss	1.5715
Initial training loss	3.0780
Training time	~41 minutes (Kaggle GPU)
LoRA rank (r)	[fill from your script]
LoRA alpha	[fill from your script]
Target modules	[fill from your script, e.g. q_proj, v_proj]
Learning rate	[fill from your script]

Training Data

Synthetic Sugar reflection dialogues generated across Gibbs, Kolb, and 5Rs frameworks. Each example consists of a multi-turn conversation where the AI only asks questions and never provides answers to the learner.

Dataset size: [X] examples

Usage