QuantumLearningMachines/qlm-math-tutor

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kPublished:May 17, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Gated Warm

The QuantumLearningMachines/qlm-math-tutor is a Llama 3.1 8B Instruct model fine-tuned with LoRA by Quantum Learning Machines to function as a Socratic math tutor for K-12 students. This model is specifically designed to never provide direct answers, instead asking guiding questions to help students independently reason through mathematical problems. It achieves a 100% Socratic question rate and 96% answer avoidance rate, making it highly effective for educational applications focused on conceptual understanding. Its primary use is to facilitate self-discovery learning in K-12 mathematics.

Loading preview...

QLM Socratic Math Tutor

The QuantumLearningMachines/qlm-math-tutor is a specialized Llama 3.1 8B Instruct model, fine-tuned using LoRA by Quantum Learning Machines. Its core purpose is to act as a Socratic math tutor for K-12 students, focusing on guiding questions rather than direct answers to foster independent problem-solving.

Key Capabilities & Performance

  • Socratic Tutoring: Achieves a 100% Socratic question rate, consistently prompting students with guiding questions.
  • Answer Avoidance: Maintains a 96% answer avoidance rate, ensuring students discover solutions themselves.
  • Relevance: 74.5% of questions are relevant to the student's specific error, with the remainder being relevant but generic guiding questions.
  • Grade-Appropriate Language: Delivers 100% of responses in language suitable for K-12 students.
  • Robust Evaluation: Metrics are rigorously evaluated with heuristic scoring (not LLM-as-judge) under production conditions, including mission context and misconception targeting.

How It Works

The model is trained to respond to student errors by asking questions that lead them to self-correction. For example, if a student incorrectly adds fractions, the model might ask, "If you had 1/3 of a pizza and 1/4 of the same pizza, would you really have less than 1/3 of a pizza total? Try drawing both fractions on the same circle."

Limitations

  • Synthetic Training Data: Trained on synthetic interactions, leading to 28% specific error targeting and 68% generic guiding questions.
  • Answer Leak Rate: Has a 1% answer leak rate, which is mitigated by a production-level filter.
  • Subject Specificity: Exclusively trained and tested for K-12 mathematics; performance in other STEM subjects is unvalidated.
  • No Longitudinal Validation: Current benchmarks measure response quality, not long-term learning gains in a classroom setting.

Good For

  • Educational platforms requiring an AI tutor that promotes critical thinking and self-discovery in K-12 math.
  • Developers looking to integrate a non-directive, Socratic teaching method into their applications.
  • Research into AI-driven educational methodologies, particularly Socratic learning.