Experiment Dashboard

A bird's-eye view of every stage. Numbers update as you progress through the pipeline.

The dashboard is empty until you load a corpus and run preprocessing. Start here →
Corpus characters
Total tokens
Vocabulary
Sentences
Train tokens
Validation tokens
Test tokens

LM1 Backoff

Test perplexity (lower is better)

LM2 Interpolation

Test perplexity (lower is better)

Best LM2 hyperparams

Not yet tuned. Run sweep →

Top tokens

Most frequent unigrams in the training split.

No data yet.

Dataset split

Tokens per split.

Not split yet.

Perplexity comparison

Test-set perplexity. Lower is better. LM1 with no smoothing can go to infinity when an unseen 4-gram appears.

No evaluation yet.