build-small-hackathon/pocket-tutor-minicpmv-socratic

Pocket Tutor Socratic LoRA

Pocket Tutor Socratic LoRA is a QLoRA adapter trained for the Pocket Tutor Space. The production app uses OpenBMB MiniCPM-V-4.6 for typed questions, worksheet images, and transcribed microphone input through one dedicated vision-language runtime.

Intended Use

• Explaining homework from photos, typed questions, or microphone input
• Producing concise Socratic hints and step-by-step support
• Helping parents ask better guiding questions
• Refusing requests to cheat on active tests or exams

Output Format

The adapter is trained on the current production UI format:

=== PROBLEM READ ===
=== KNOWNS ===
=== STRATEGY ===
=== WORKED STEPS ===
=== CHECK ===
=== NEXT HINT ===
=== PARENT NOTE ===

Training Recipe

• Base model: openbmb/MiniCPM-V-4.6
• Method: QLoRA with 4-bit NF4 quantization
• Hardware: Modal NVIDIA A10G
• Training data: 16 structured tutoring examples plus 7 schema-aligned follow-up turns
• Sequence length: 1536 tokens
• Runtime pairing: the same MiniCPM-V base model used by the Pocket Tutor Space

Validation

The adapter is smoke-tested on Modal against every structured training example and follow-up turn. The smoke test fails if any of the seven production UI sections fall back to default placeholder text.

Limitations

This model can make mistakes and should not be used to cheat on graded work. It is designed to teach process and reasoning, not replace a teacher.