mjpsm/activity-generation-model-v0.2
mjpsm/activity-generation-model-v0.2 is a 0.5 billion parameter instruction-tuned Qwen2.5 model fine-tuned by mjpsm to generate simple next learning activities from student knowledge submissions. With a 32768 token context length, it specializes in proposing small, actionable follow-up activities, outputting structured JSON with a title, description, instructions, and one of seven activity types. This model is designed for adaptive learning workflows, directly building on prior student learning.
Loading preview...
Activity Generation Model v0.2
mjpsm/activity-generation-model-v0.2 is a fine-tuned version of Qwen/Qwen2.5-0.5B-Instruct specifically designed to generate a single, simple next learning activity based on a student's previous knowledge submission. This model is part of the MyVillage Project's activity-generation workflow, aiming to propose logical next steps that build directly on what a student has learned, practiced, created, researched, or experienced.
Key Capabilities
- Structured Activity Output: Generates activities in a JSON format, including
title,description,instructions, andactivityType. - Defined Activity Types: Supports seven distinct activity categories:
REFLECTION,RESEARCH,COLLABORATE,CREATE,PRACTICE,EXPERIENCE, andTEACH. - Small, Actionable Steps: Activities are designed to be promptly completable, provide clear instructions, and avoid large, multi-part assignments.
- Direct Relevance: Activities directly build on the student's previous knowledge submission.
- Fine-tuned Performance: The model was fine-tuned using LoRA with a focus on achieving the lowest validation loss, ensuring optimized generation for its specific task.
Good For
- Adaptive Learning Systems: Proposing follow-up activities in response to student progress.
- Educational Content Generation: Creating structured activity data for downstream application logic.
- Extending Prior Learning: Generating next steps that logically advance a student's demonstrated knowledge.
Limitations
It's important to note that the model's training includes synthetically generated activities, and token-level accuracy does not directly measure pedagogical correctness. Generated activities should be reviewed for relevance and appropriateness, especially in higher-stakes educational settings. The model is intended for single next activities, not long-term learning plans, and performance outside its training distribution is not fully characterized.