kubeflow / sdk / 41143660549

85%

push

github

fix(trainer): validate TorchTune loss and dataset source enums (#742)

TorchTuneConfig.loss was never type-checked, and the ValueError guarding
TorchTuneInstructDataset.source dereferenced `.value` on the very object it
had just established was not a DataFormat. Both paths therefore surfaced an
AttributeError from inside the SDK instead of naming the offending field,
even though `dtype` in the same function already raised a clear ValueError.

Passing a bare string is an easy mistake when a fine-tuning configuration is
written by hand, and DataFormat.JSON.value is literally "json", so the wrong
value tends to look correct at a glance. Reporting it where the config is
converted to trainer args keeps the diagnosis next to the cause.

Assisted-by: Claude Code

Signed-off-by: Yigtwxx <yigiterdogan023@gmail.com>

3 of 3 new or added lines in 1 file covered. (100.0%)

6324 of 7441 relevant lines covered (84.99%)

1.7 hits per line