eth-nlped/Eduardo-27b

Overview

Eduardo-27B is a 27 billion parameter math tutor model, fine-tuned from Qwen/Qwen3.8-27B by eth-nlped using multi-turn reinforcement learning (DPPO). Its core purpose is to teach students how to solve math problems through interactive, Socratic dialogue, rather than simply providing answers. The model was trained in a simulated classroom environment, engaging in multi-turn conversations (up to 22 turns) with a frozen Llama-3.1-8B-Instruct student on Big-Math problems. The training reward mechanism focuses on the student's normalized learning gain on a masked near-transfer post-test, ensuring the tutor guides understanding without directly handing over solutions.

Key Capabilities

• Socratic Math Tutoring: Emphasizes guiding questions and prompting for justification, significantly reducing direct explanations and answer provision compared to its base model.
• High Pedagogical Alignment: Achieves the highest pedagogical reward model (Ped-RM) score and win rate against human teachers on MathDial, outperforming frontier models like Gemini 3.1 Pro.
• Efficiency: Uses 2.4–6.2 times fewer thinking tokens per turn than frontier models on TutorMoments, indicating efficient reasoning.
• Avoids Over-help: Demonstrates strong performance in avoiding over-help, ensuring students do the cognitive work.

Good For

• Interactive, Step-by-Step Math Tutoring: Ideal for applications requiring an AI to guide users through mathematical problem-solving.
• Socratic and Scaffolded Dialog Systems: Suitable for developing conversational agents that encourage critical thinking and self-discovery.
• Research on Pedagogical LLM Alignment: A valuable research artifact for studying how LLMs can be trained to be effective educators.