Get startedModels · Optimization
Train and fine-tune models
Train from reviewed boxes or object outlines. Detection candidate search uses a configured engine; segmentation and experimental reward learning run locally.
Source reviewed 2026-09-16. Availability depends on your installation, permissions, and compatible runtimes. A supported path is not a guarantee of model quality or production readiness.
Outcome
Create an inspectable model artifact and distinguish successful export from measured improvement.
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
- — An immutable dataset version
- — Architecture or foundation, hyperparameters, seed, and compute target
Expected output and limits
- — Checkpoints and a registered model version
- — Logs, curves, configuration, spend, and lineage
Core workflow
Choose data and strategy
Select reviewed detection or segmentation data. Prepare training split if needed; it creates a new version without confirming unreviewed labels. Clips contribute sampled frames.
Select compute
Choose This machine for local training. Smart and Pro can connect a supported GPU provider with an API key and select that account for compatible remote detection runs. Score Studio coordinates the job; the provider supplies and charges for the GPU. Segmentation and reward optimization currently run locally.
Observe optimization
Track status, loss curves, validation metrics, logs, cost, and early stopping.
Verify the result
Treat artifact readiness and measured improvement as separate outcomes; compare the new version on a compatible immutable benchmark.
What this surface supports
Supervised detection and local segmentation
Configured automated detection search
Experimental vision-weight reward optimization
Training and validation telemetry
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
- — Training/validation loss gap
- — Task metrics by class and slice
- — Learning rate and throughput
- — Early-stopping criterion
Common failure modes
- — Overfitting
- — Divergent loss or invalid labels
- — Out-of-memory compute
- — Non-reproducible configuration
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
- — Runs continue in the background
- — Use only the controls supported by that run
- — Artifact export never implies measured improvement