Open up a training workflow: synthetic images, bounded settings, explicit optimization and held-out evaluation.
Scope: original public experiment. I built the application workflow and interface.
Use it to explore:
Explain the boundary between a model's recommendation and an actual computation.
Compare fixed training settings on a deliberately small, reproducible task.
Bound settings
Generate data
Train
Evaluate
What happens inside
The demo generates 80 training and 40 test images, each 12×12 pixels. NumPy performs logistic-regression updates. Server-Sent Events carry progress to the browser.
Decisions and tradeoffs
Keep the computation inspectable
A small logistic-regression model makes the update loop visible. The task is intentionally simple; a high score on circles and squares says little about real images.
Constrain the plan
Learning rate is bounded to 0.05–3.0 and epochs to 10–60. The model can suggest settings, but the application controls their limits.
Separate suggestion from measurement
Fallback settings are labeled as defaults. Training metrics are computed by code; a generated summary is a separate interpretation.
Recorded example
Recorded deterministic baseline
Explore the actual image samples, weight map and metrics below. Learning rate 1.0; 40 epochs; fixed data seeds. No provider request is needed.
What this does not establish
This is a synthetic classification experiment, not large-scale model training. The recorded baseline uses fixed settings, not an agent-selected improvement. Its held-out result does not establish real-world generalization.
Try a live run
Live requests send your prompt to OpenRouter and the selected model provider. Use public, non-sensitive material. Read how demo data is handled.
Checking live availability…
Technical walkthrough reviewed September 2026. Examples do not imply client adoption or production maturity.