VertexAI Tabular AutoML rejecting rows containing nulls

I am trying to build a binary classifier based on a tabular dataset that is rather sparse, but training is failing with the following message:

Training pipeline failed with error message: Too few input rows passed validation. Of 1169548 inputs, 194 were valid. At least 50% of rows must pass validation.

My understanding was that tabular AutoML should be able to handle Null values, so I'm not sure what's happening here, and I would appreciate any suggestions. The documentation explicitly mentions reviewing each column's nullability, but I don't see any way to set or check a column's nullability on the dataset tab (perhaps the documentation is out of date?). Additionally, the documentation explicitly mentions that missing values are treated as null, which is how I've set up my CSV. The documentation for numeric however does not explicitly list support for missing values, just NaN and inf.

The dataset is 1 million rows, 34 columns, and only 189 rows are null-free. My most sparse column has data in 5,000 unique rows, with the next rarest having data in 72k and 274k rows respectively. Columns are a mix of categorical and numeric, with only a handful of columns without nulls.

The data is stored as a CSV, and the Dataset import seems to run without issue. Generate statistics ran on the dataset, but for some reason the missing % column failed to populate. What might be the best way to address this? I'm not sure if this is a case where I need to change my null representation in the CSV, change some dataset/training setting, or if its an AutoML bug (less likely). Thanks!

To allow invalid & null values during training & prediction, we have to explicitly set the allow invalid values flag to Yes during training as shown in the image below. You can find this setting under model training settings on the dataset page. The flag has to be set on a column by column basis.