Reference#
What a run is made of, one page per part. Start with Configurations if you are looking for a key, and with Tasks if you are looking for what a run does with one. These pages assume a run has already been made, which Quickstart covers, and they are what Writing your own method points at when a method of your own needs one of them.
Running an experiment#
Configurations is every registration in the library, as searchable cards built from the registry itself.
Tasks is what one experiment is: the stages it runs, and what it writes to disk.
Reproducing published results is a grid of tasks, and how a published paper’s values are registered beside the library rather than inside it.
Configuring a model#
Token vectors is what a model starts from, whether a pretrained vector file or one-hot inputs.
Metrics is what a run scores, and what a class-imbalanced corpus needs.
Highlight supervision is training against a highlight annotation rather than against the label alone.
Backbones is the encoder contract and the three implementations that satisfy it.
Reading a run back#
Analyzers turns a results directory into a frame, a table, or the text a model selected.
Faithfulness is sufficiency and comprehensiveness, and what each compares.
Diagnostics records the inside of a forward pass without changing what it computes.
Computational cost reports what a run spent.
Interlocking and what the bottleneck guarantees is what the bottleneck does and does not guarantee, measured.
Corpora#
Datasets is the corpora, their loaders, and the preprocessors that prepare them.
Data is the batch contract every loader and model is written against.