Build a corpus
Import permitted text, track provenance and separate training material from held-out evaluation.
Train a persistent language model from an approved corpus. Explore its learning, build focused agents and inspect the evidence behind their results.
Start locally in your browser. No account needed for local use.
Import permitted text, track provenance and separate training material from held-out evaluation.
Update word-pair counts and neural character weights in bounded batches. Inspect model revisions, loss and exportable checkpoints.
Save reusable evidence retrieval, word-prediction and neural-character agents with explicitly limited capabilities.
Local training stays on your device. Hosted training requires installation administrator access and uses separate private storage. Projects, analytics, credentials and enterprise inputs are not silently collected into the corpus.
This implementation is a word-bigram language model and a compact 890-parameter neural character predictor, not a general-purpose conversational LLM. Agents return source excerpts or learned prediction probabilities. Corpus quality determines usefulness.
Training and data policy →