MLflow evaluate failed with error 'Positional indexers are out-of-bounds' on validation dataset

TL;DR

The validation dataframe's index is being used as row positions somewhere it cannot fit: reset the index on the dataset before evaluate, and confirm the predictions and targets align by position, not by a filtered index left over from preprocessing. A filtered dataframe keeps its original index labels, and positional indexing against it goes out of bounds on the gaps.

MLflow evaluate failed with error 'Positional indexers are out-of-bounds' on validation dataset

Use this when

• The run or call fails with this symptom: Positional indexers are out-of-bounds from MLflow evaluate on a validation dataset.
• You are working on mlflow.evaluate on a validation set produced by filtered preprocessing and the failure repeats on the same run or configuration.
• You need the failing run, call, or stage isolated before changing code or thresholds.

Not for this skill when

• A genuinely empty validation frame after filtering, where the fix is the filter rather than the index.
• The error text names a different tool, stage, or root cause. Match the verbatim error first and use that skill.

Steps

1. Inspect the validation dataframe's index: its length, and whether labels are contiguous from zero. Verify: the index state is known
2. Reset the index with drop set so predictions, targets, and features align by position. Verify: the frame has a clean positional index
3. Confirm predictions were generated from this exact frame in this order, not from an earlier unfiltered version. Verify: prediction count equals frame length
4. Rerun evaluate and confirm metrics and plots generate without the indexer error. Verify: evaluation completes with a metrics table

Variant phrasings

mlflow evaluate positional indexers out of bounds

The verbatim phrasing.

pandas filtered index mlflow evaluation error

The filtered-index phrasing.

mlflow evaluate validation dataset index reset fix

The reset phrasing.

Compatibility: MLflow 2.x and 3.x evaluate API with pandas validation data

Why it happens

Pandas preserves original index labels through filtering, so a frame can hold 800 rows labelled up to 5000. Positional operations inside evaluation assume dense positions, and the first gap beyond the frame's length raises the out-of-bounds error far from the filtering step that caused it.

Edge cases / pitfalls

• Custom metrics functions receive the frame too; a reset index fixes their positional bugs as well.
• If predictions come from a model server, regenerate them after filtering rather than slicing an older prediction array.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_g4GQwNUxCdum9lJ02vAh-Q