mjpsm/activity-generation-model-v1
The mjpsm/activity-generation-model-v1 is a 0.5 billion parameter causal language model, fine-tuned from Qwen2.5-0.5B-Instruct by mjpsm. It specializes in generating structured educational activities (title, description, instructions) based on a learner's village goal, previous activity, and knowledge submission. This model is optimized for creating personalized, next-step learning activities within adaptive learning environments, outputting responses in a strict JSON format.
Loading preview...
Overview
The mjpsm/activity-generation-model-v1 is a 0.5 billion parameter causal language model, fine-tuned from Qwen/Qwen2.5-0.5B-Instruct using Supervised Fine-Tuning with LoRA. Its core function is to generate a structured educational activity, including a title, description, and instructions, based on three key inputs: a learner's village goal, their previous activity, and their knowledge submission. The model was developed for the MyVillage learning environment to facilitate personalized learning progressions.
Key Capabilities
- Contextual Activity Generation: Creates relevant next activities by considering a learner's broader goal and immediate learning context.
- Structured JSON Output: Guarantees output in a predefined JSON schema, making it easy to integrate into applications.
- Educational Focus: Specifically designed to build on demonstrated knowledge and guide learners toward larger objectives.
- Small Footprint: At 0.5B parameters, it offers efficient deployment for its specialized task.
Intended Use Cases
This model is ideal for experimental educational activity generation, particularly in adaptive learning systems. It can be used for:
- Generating personalized learning activities.
- Building learning progression and recommendation systems.
- Prototyping adaptive learning environments where activities are dynamically created based on learner input.
Limitations
As an experimental model, it may occasionally generate activities that are too broad or narrow, not perfectly aligned with goals, or produce invalid JSON. It was trained on synthetic data, so real-world performance should be independently evaluated. It does not predict activity type, duration, or mastery scores.