Model workspace.
Build knowledge, train a model and assemble agents in one workspace.
Documents and model state stay in this browser. Browser storage is not encrypted or backed up automatically.
Corpus library
Start with the included teaching corpus or import text you have permission to train on.
No documents loaded.
Train your model
Each batch updates word-level statistics and a compact neural character predictor. Inspect both models and export their checkpoints.
Enabled local training continues while this workspace is open and visible.
Model exports can reveal learned training text. Keep private. Deleting a learned document clears the model and queues the remaining corpus for rebuilding.
Create an agent
Configure a reusable, read-only agent that uses the selected model.
Results will show excerpts with provenance or learned prediction probabilities. Queries are not added to training.
Evaluate and verify
Held-out perplexity measures next-token prediction on documents not trained into the model.
Add separate evaluation documents. Lower perplexity is better only for comparable data and vocabulary; it is not a measure of factual accuracy. Neural evaluation uses up to 2,048 held-out characters; independent word and character scores are not directly comparable.
Corpus ingestion beyond this workspace
Operators can import permitted text files and explicitly allowlisted HTTPS plain-text URLs using the bundled private CLI. Hosted batch training can run through optional cron. No arbitrary web crawler runs on the public server.
Read the training guide ↗Only approved corpus content is learned. Ordinary CLARA projects, enterprise data and your agent queries are not automatically collected. Data policy.