Understand VLM training availability

Multimodal dataset schemas do not imply an executable vision-language model trainer. General VLM fine-tuning is not an implemented training path in this release.

Source reviewed 2026-09-23. Availability depends on your installation, permissions, and compatible runtimes. A supported path is not a guarantee of model quality or production readiness.

Outcome

Choose a supported detection or local segmentation training path, or export compatible annotations for an external training system.

Start here

What you need to know first

Check the inputs below before starting. If you are new, begin with the first-project guide. A dataset holds media and labels; a model produces results; a deployment makes a selected model version callable. Creating one does not create the others.

Bring these inputs

Your intended task and annotation schema
An external training system if you need VLM fine-tuning

Expected output and limits

No in-app VLM training artifact is promised
Format-dependent data export for an external system

Core workflow

01
Check the task
The in-app trainer supports detection and local instance segmentation. A VQA or captioning schema is a data format, not proof of training support.
02
Prepare portable data
Review your images or sampled frames and annotations. Check that the export format preserves the fields your external trainer needs.
03
Choose a supported route
Use Training for supported detection or segmentation data. Do not launch a detector expecting it to learn image-question answering.
04
Verify external results
External models need a compatible runtime and artifact before Score Studio can run them. External training is not automatically registered or evaluated.

What this surface supports

Multimodal annotation schemas
Format-dependent annotation export
Detection training
Local segmentation training

Expert section

Contracts, signals, and failure modes

Use this section when you are defining acceptance criteria, automating the surface, or reviewing whether its output is safe to promote downstream.

Quality and operating signals

Task accuracy or answer quality
Grounding IoU/recall
Hallucination and abstention rate
Validation loss

Common failure modes

Catastrophic forgetting
Prompt-template leakage
Training on answers without visual evidence
Evaluating on near-duplicate images

Expert release checklist

Inputs and dependencies are pinned to immutable versions.
Acceptance metrics include critical classes and operating slices.
Failure, retry, cost, and rollback behavior are understood.
The resulting artifact has an owner and a downstream review path.

Engineering safeguards

No general VLM trainer is promised
Verify export fields before external use
Runtime compatibility is required