# Inconsistent predictions when using custom ML backend with overlapping labels

**URL:** <https://community.labelstud.io/t/inconsistent-predictions-when-using-custom-ml-backend-with-overlapping-labels/585>\
**Category:** Machine Learning\
**Created:** [June 27, 2025, 2:52pm UTC](https://community.labelstud.io/t/inconsistent-predictions-when-using-custom-ml-backend-with-overlapping-labels/585 "2025-06-27T14:52:52Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![kimdyn](https://avatars.discourse-cdn.com/v4/letter/k/e5b9ba/32.png) [@kimdyn](https://community.labelstud.io/u/kimdyn)\
**Post date:** [June 27, 2025, 2:52pm UTC](https://community.labelstud.io/t/inconsistent-predictions-when-using-custom-ml-backend-with-overlapping-labels/585/1 "2025-06-27T14:52:52Z")

</div>

Hi all!

I am integrating a custom ML backend with Label Studio to auto-suggest labels during annotation. 🙂It works well for simple single-label tasks; but when using overlapping labels (multi-label classification); the predictions are inconsistent sometimes only one label shows up / confidence scores are missing entirely. 🤔

I suspect the issue might be with how the backend formats multiple predicted labels or with how Label Studio interprets overlapping annotations from the model. 😇 I followed the ML backend example, but the docs don’t cover multi-label /overlapping scenarios in depth. Checked [Label Studio Documentation — Write your own ML backend](https://labelstud.io/guide/ml_create) with [JIRA Course](https://www.igmguru.com/agile-and-scrum/jira-project-management-training) documentation guide for reference.

Has anyone successfully set up a backend that returns multiple predicted regions or tags for the same object/text span? 🤔 Would love a working example or advice on the right payload format to ensure consistency in pre-labeling.

Thank you !! 🙂