uf-aice-lab/llama3-rlrgd-edu-trustworthy

---

language: en license: apache-2.0 tags: - education - trustworthy-ai - reinforcement-learning - llama-3 - large-language-model - alignment - k12 - ai-education datasets: - custom-k12-student-teacher-dialogues - safe-generated-reverse-data model-index: - name: Llama 3 - RLRGD Educational Trustworthy AI results: - task: text-generation dataset: custom-k12-student-teacher-dialogues metrics: - name: EvaluLLM (Edu-Facilitativeness) type: custom value: 0.87 - name: BLEU type: bleu value: 21.3 - name: ROUGE-L type: rouge value: 43.8

📘 Llama 3 - RLRGD Educational Trustworthy AI

This model is a fine-tuned version of Meta's Llama 3, developed using our novel RLRGD (Reinforcement Learning from Reverse-Generated Data) framework to meet the demands of trustworthy AI in K-12 education. It is designed to provide pedagogically appropriate, safe, and engaging responses in the context of educational dialogue—particularly in math education scenarios involving student-teacher Q&A.

🧠 Model Details

• Base Model: Llama 3 (13B)
• Training Method: RLRGD (custom reinforcement learning pipeline with contrastive learning from reverse-generated data)
• Data Sources:
  • Real K-12 student-teacher dialogues from the ALTER-Math dataset (anonymized and preprocessed)
  • Reverse-generated unsafe/inadequate samples for contrastive feedback
  • Safe-LLMs generated synthetic augmentations
• RL Objective: Encourage the model to maximize alignment with teacher-like facilitative responses and minimize unsafe, irrelevant, or overly verbose outputs

✅ Intended Use

This model is intended for use in AI-powered educational applications, such as:

• Interactive math tutors
• AI-powered teacher assistants
• Educational chatbot environments for K-12 students

It is ideal for scenarios requiring trustworthiness, personalization, and pedagogy-aware feedback.

❗ Limitations

• Although the model has been fine-tuned for safety and educational soundness, it should not be used without teacher supervision in high-stakes learning environments.
• The model may still produce hallucinated or out-of-scope answers when prompted with out-of-domain queries.
• Performance may vary outside of math education or K-12 contexts.

🛡️ Safety and Alignment

The model was trained with safety and alignment in mind:

• Reverse-generated contrastive training to suppress unsafe or unhelpful outputs
• Human-in-the-loop reinforcement using expert-rated teacher dialogues
• Use of Safe-LLMs during generation and evaluation for toxicity filtering

We apply strict preprocessing:

• Removal of personal identifiers
• Filtering of inappropriate language
• Scaffolding principles from ZPD, Constructivism, and Scaffolding Theory