art87able/unstuck-qwen2.5-0.5b-steps
The art87able/unstuck-qwen2.5-0.5b-steps model is a 0.5 billion parameter Qwen2.5-Instruct-based model fine-tuned by art87able. It specializes in breaking down overwhelming tasks into tiny, timed, and categorized steps, outputting them in a specific JSON schema. This model is optimized for ADHD task management, providing concrete actions with estimated minutes and categories, and is designed for fully local, serverless-free operation.
Loading preview...
Overview
art87able/unstuck-qwen2.5-0.5b-steps is a compact 0.5 billion parameter instruction-tuned model based on Qwen/Qwen2.5-0.5B-Instruct. Developed by art87able, its core function is to transform a single, overwhelming task into a series of small, manageable, and time-estimated steps, formatted as a precise JSON output. This model was specifically fine-tuned for the Unstuck application, which focuses on ADHD task breakdown.
Key Capabilities
- Task Decomposition: Breaks down complex tasks into individual, concrete actions starting with an imperative verb.
- Structured Output: Generates output strictly adhering to a predefined JSON schema, including
text,category(admin, creative, errand, deep-work), andest_minutes(positive integer, never above 25). - Efficiency: Designed for local, serverless-free deployment, powering the
UNSTUCK_BACKEND=finetunedpath within the Unstuck application.
Training Details
The model was trained using LoRA (r=16, α=32, dropout 0.05) over 3 epochs with a learning rate of 2e-4. The training data consisted of 130 schema-valid task breakdowns, synthetically distilled from a larger serverless model (Qwen/Qwen3-30B-A3B-Instruct-2507) and rigorously validated against Unstuck's schema.
Limitations
This is a highly specialized model, not intended for general chat or broad language understanding. Its outputs should ideally be schema-validated, and minute estimates serve as starting points, as the Unstuck application recalibrates timings based on user data.