typedecide.errors¶
Every exception the library raises deliberately. The error-handling guide says which module raises which class and what the CLI does with them.
Every error this library raises, so a caller can catch what it means to catch.
Messages name the thing that is wrong and what would fix it. A stack trace is not a diagnosis, and someone running a twelve-hour fine-tune deserves better than one.
Bases: Exception
Base class. Catch this to catch anything this library raises deliberately.
Bases: TypeDecideError
A decision, criterion or option is not well formed.
Bases: TypeDecideError
A dataset could not be read, or is unusable as given.
Bases: TypeDecideError
The prompt and the tokenizer disagree, so a readout would be silently wrong.
Bases: TypeDecideError
Training cannot start, or cannot be trusted if it did.
Bases: TypeDecideError
A model could not be turned into the artefact that was asked for.
Bases: TypeDecideError
A configuration file or CLI invocation does not describe a runnable job.
Bases: TypeDecideError
An evaluation started from usable inputs and still could not produce a score.
Distinct from ConfigError (the job was described wrongly) and DataError (the
dataset is unusable): this is what goes wrong while scoring -- most often a
scorer that breaks its contract by returning the wrong number of probability
vectors, a vector of the wrong width, or values that are not probabilities.
Bases: EvaluationError, ConfigError
A ScoreFn broke its contract: wrong shape, or values that are not probabilities.
Catch EvaluationError. This class also inherits ConfigError purely for backward
compatibility: 0.1.0 raised ConfigError for a misbehaving scorer, so
except ConfigError written against that release keeps working. The ConfigError
base is deprecated and will be dropped in a future major version.