Encord × Weights & Biases: How to Keep Your Training Data in Lockstep with Ground Truth

We’re excited to announce a native integration between Encord and Weights & Biases that syncs your latest annotation data across both platforms - no manual exports, SDK scripting, or data engineering required. Your training datasets in Weights & Biases will stay continuously aligned with the ground truth labels your team creates in Encord.

What the integration does

With a few clicks, you can connect an Encord project to a Weights & Biases workspace. From that moment on, annotation updates approved in Encord are pushed to Weights & Biases as Artifacts, so anything you train, evaluate, or track in Weights & Biases reflects the most current labels and metadata.

• One-time connect: Authenticate Encord → Weights & Biases and choose where your data should land.
• Auto-sync: New or edited labels in Encord flow into Weights & Biases as versioned Artifacts.
• Ready for pipelines: Point Weights & Biases runs and sweeps at the synced Artifacts - no extra plumbing.

Why it matters

• Eliminates manual handoffs: No more exporting, transforming, or re-uploading datasets between tools.
• Keeps training data fresh: Model iterations always train on up-to-date ground truth, reducing label drift and stale experiments.
• Bridges data and model workflows: Encord is the source of truth for labeled data; Weights & Biases remains the system of record for training and experiment management.
• Reduces engineering effort: A managed, click-to-connect data flow means ML teams can ship faster without custom glue code.

How it works

1. Connect your Encord project to a Weights & Biases workspace/project.
2. Select the datasets and label schema you want to sync.
3. Sync happens automatically: as annotations are created or updated in Encord, we package and version them, then push to Weights & Biases as Artifacts (and accompanying metadata).
4. Train & track in Weights & Biases using those Artifacts. Your dashboards, comparisons, and regressions all reflect the latest labels - by default.

Under the hood, Encord monitors annotation states and handles the heavy lifting (packaging, versioning, and delivery) so your team doesn’t have to.

For teams using CoreWeave, the integration also supports native CoreWeave AI Object Storage ingestion, making it easy to scale data flows without custom infrastructure.

Learn how to create a CoreWeave integration in Encord.

Watch the full integration demo:

Key use cases

• Accelerated model iteration: Curate and annotate in Encord; as soon as labels are reviewed, Weights & Biases receives updated artifacts. Your next run pulls the freshest data - no “are we training on the latest labels?” guesswork.
• Data-centric debugging: When model performance dips, jump into Weights & Biases to trace the exact data and labels used. Identify failure modes, then target fixes in Encord (e.g., adding edge cases or correcting labels). The updates auto-sync back to Weights & Biases for the next experiment.
• Operational visibility: Product managers, data scientists, and MLOps engineers can rely on Weights & Biases dashboards to reflect current dataset status and experiment outcomes without chasing file drops or ad-hoc scripts.
• GenAI & human evaluations: Manage human feedback and rubric-based evaluations in Encord and log results to Weights & Biases for side‑by‑side analysis so qualitative signals sit alongside metrics and loss curves.

Get Early Access

Interested in being among the first to use this integration? We're currently offering early access to select teams. Sign up here to get early access and see how Encord’s integration with Weights & Biases can improve synchronization across your workflow.