> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/sdgm/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/sdgm/_mcp/server.

# weight_mode

#### `weight_mode: (sample | weighted_loss)` (Optional)

## Description

`weight_mode` specifies how Kumo applies training weights when either a training-table `WEIGHT` column is present and/or `majority_sampling_ratio` is configured for binary classification.

The training-table `WEIGHT` column can come from either:

* SDK/custom training-table workflows (see [How do I assign different weights to training samples?](/troubleshooting/can-i-perform-weighted-training))

* generated weights via [weight\_col](/reference/weight_col) in `training_table_generation`

* `sample`: samples training examples with replacement according to the effective weight distribution.

* `weighted_loss`: weighs training examples in the loss function according to the effective weight distribution.

When both a training-table weight column and `majority_sampling_ratio` are set for binary classification:

* In `sample` mode, row and class weights are combined for sampling.
* In `weighted_loss` mode, row and class weights are combined in loss weighting.

### Supported Task Types

* All

### Default Values

When `weight_col` is configured, Kumo defaults to `weighted_loss` -- the AutoML plan sets `weight_mode: [weighted_loss]` automatically, so the loss function is weighted by the column values without further AutoML search. When only `majority_sampling_ratio` is set for binary classification (no `weight_col`), AutoML does not set a default `weight_mode` and you must choose one explicitly.