Model Card for Llama-3.2-3B-SCRIBE

Abstract

Language models can be used to provide interactive, personalized student feedback in educational settings. However, real-world deployment faces three key challenges: privacy concerns, limited computational resources, and the need for pedagogically valid responses. These constraints require small, open-source models that can run locally and reliably ground their outputs in correct information. We introduce SCRIBE, a framework for multi-hop, tool-augmented reasoning designed to generate valid responses to student questions about feedback reports. SCRIBE combines domain-specific tools with a self-reflective inference pipeline that supports iterative reasoning, tool use, and error recovery. We distil these capabilities into 3B and 8B models via two-stage LoRA fine-tuning on synthetic GPT-4o-generated data. Evaluation with a human-aligned GPT-Judge and a user study with 108 students shows that 8B-SCRIBE models achieve comparable or superior quality to much larger models in key dimensions such as relevance and actionability, while being perceived on par with GPT-4o and Llama-3.3 70B by students. These findings demonstrate the viability of SCRIBE for low-resource, privacy-sensitive educational applications.

---

Model Description

Llama-3.2-3B-SCRIBE is a fine-tuned large language model for interactive educational feedback.
It implements the SCRIBE framework: structured chain reasoning with multi-hop tool calling and self-reflection, enabling small models to deliver pedagogically valid, actionable, and context-grounded explanations to student questions.

• Developed by: EPFL (Machine Learning for Education Lab)
• Paper: SCRIBE: Structured Chain Reasoning for Interactive Behavior Explanations using Tool Calling
• Authors: Fares Fawzi, Vinitra Swamy, Dominik Glandorf, Tanya Nazaretsky, Tanja Käser
• Model type: Llama-3.2 3B fine-tuned with two-stage LoRA
• Languages: English
• License: Apache 2.0
• Finetuned from: Llama-3.2 3B base

---

Uses

Direct Use

The model is designed to:

• Provide personalized, interactive feedback to students in MOOCs.
• Generate pedagogically grounded explanations using multi-step reasoning.
• Support privacy-sensitive deployments by running locally on very modest hardware compared to 8B and larger models.

Notes

• The 3B variant is optimized for resource-constrained environments, making it well-suited for laptops, personal servers, or classroom settings where GPU access is limited.
• Compared to the 8B model, the 3B achieves lower absolute performance but still benefits from the SCRIBE framework’s improvements over vanilla base models.

---

Training Details

Training Data

• Base data: Feedback reports from MOOCs (DSP, GEO, VA, LNV)
• Synthetic data: ~7,000 student-like questions generated with GPT-4o, including reasoning traces, tool calls, and final responses
• Real data: 75 student-authored questions annotated into pedagogical categories

Training Procedure

• Two-stage LoRA fine-tuning:
  1. Stage 1: Initial reasoning + tool selection
  2. Stage 2: Multi-hop reasoning + final answer generation
• Inference: Closed-loop tool-calling with self-reflection and error recovery

Training Hyperparameters

• Regime: bf16 mixed precision
• LoRA rank: 128 for 3B variant

---

Citation

If you use this model, please cite:

APA
Fawzi, F., Swamy, V., Glandorf, D., Nazaretsky, T., & Käser, T. (2025).
SCRIBE: Structured Chain Reasoning for Interactive Behavior Explanations using Tool Calling. EPFL.

BibTeX

@inproceedings{2025-EMNLP-Scribe,
  author    = {Fares Fawzi and Vinitra Swamy and Dominik Glandorf and Tanya Nazaretsky and Tanja K{\"a}ser},
  booktitle = {Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  title     = {SCRIBE: Structured Chain Reasoning for Interactive Behavior Explanations using Tool Calling},
  year      = {2025}
}