MicroComp IT Demo Lab · Machine learning fundamentals
Configure the same decoder-only architecture that will power a real browser trainer, inspect how text becomes tokens, and watch real training events in your browser.
Choose where the work happens
Privacy and retention behavior stay visible before a training run begins.
Move through the model lifecycle
Each stage answers one question about how a tiny language model learns.
Stage 1 · Data
Original fictional alert summaries paired with safe incident-response descriptions. No download or user upload is required.
Spaces and common letters appear most often, so the model sees them most often during training.
Stage 2 · Tokenization
This teaching box only previews character IDs. It does not prompt, train, or generate with the model.
Unknown characters use <unk> · ID 3. To keep this visualization readable, at most 160 token chips are displayed.
alert:
space
The model receives earlier character IDs and learns to predict the character that follows.
Stage 3 · Architecture
Every input position can attend only to itself and earlier positions.
Stage 4 · Training
Downloading preserves a portable copy. Saving to this browser validates the same package and makes the model persistent for local inference.
Follow token flow and replay bounded observations captured from the real model whenever the selected engine provides them.
Waiting for a checkpoint.
Stage 5 · Prediction Lab
The model scores four complete, readable response plans for an alert it did not train on. The security-area labels describe those curated plans; they are not generated text.
Select a saved local model to begin.
| Response candidate | Baseline rank/loss | Trained rank/loss | Relative score | Result |
|---|
| Held-out alert | Model selected | Reference | Result |
|---|
Free generation is preserved as a limitations demonstration. Readable response ranking above is the recommended educational activity.
Complete a local training run to use the model here.
Remember: This tiny model learns character patterns from a small synthetic dataset. Its output is educational, not reliable cybersecurity guidance.
Completed models stay in this browser until you delete them or clear site storage. Export a portable copy for durable backup.
Loading models stored in this browser…
Stage 6 · Analysis
The model has not learned enough structure. Both training and validation loss remain high, and output looks mostly random.
Training loss continues falling while validation loss worsens. The model is fitting its lessons more than unseen examples.
A tiny dataset can be repeated instead of understood. Familiar phrases alone are not evidence of reasoning.
Useful patterns carry into held-out validation records, even when exact wording was not present during training.
Lower values favor likely characters. Higher values increase variation and usually increase mistakes in a tiny model.
More width and layers add capacity and computation. The 200,000-parameter ceiling keeps tradeoffs visible and bounded.
Complete or load a model with training history to generate a measured, plain-language report.