AI Workforce · Module
AI Lab & Data Labeling
Datasets, labelling, experiments and models, governed for consent and licence
AI Lab & Data Labeling manages datasets, labelling projects, experiments and model versions for the AI Workforce. It is eligible from W2 with an internal light version in Year 1; a Data Labeling service with vetted experts is planned for W4.
The lab behind better agents
Some agent skills need their own classifiers, extractors or adapters. AI Lab & Data Labeling is eligible from W2; in Year 1 a light internal version supports One3's own work, and the Data Labeling service is planned for W4.
Datasets with provenance
The dataset registry tracks versions, datasheets, lineage, consent tier and classification. PII is redacted before labelling and reviewer access is logged. When a data principal asks for erasure, matching dataset rows are marked and retraining or exclusion follows the policy. Licence compliance for datasets and base models is tracked against Fact Register rows.
Labelling projects
Labelling projects come from templates for classification, extraction and preference pairs, with assisted pre-labelling. Quality is managed through consensus QA, inter-annotator agreement and an adjudication queue. The planned Data Labeling Service adds a vetted expert workforce through Human Expert Cloud.
Experiments and models
Experiment tracking is linked to eval runs. The model registry integrates with the Model Hub router, with canary promotion and rollback. Fine-tune jobs for classifiers and extractors pass a licence gate, have a cloud fallback and are promoted only if they beat the prompt-only baseline. Tenant adapter training jobs are tenant-scoped and work with the Training Gym. Compute budgets apply per job, with One3 Compute GPU pools planned.
How it fits
AI Lab sits between the Training Gym, which supplies consented traces and scenarios, and Model Hub, which serves promoted models. It uses Synthetic Data Studio where real data should not be used.
Features
What AI Lab & Data Labeling includes
Dataset registry
Versions, datasheets, lineage, consent tier and classification.
Labelling projects
Classification, extraction and preference-pair templates with pre-labelling.
Experiment tracking
Linked to eval runs.
Model registry
Integrated with Model Hub routing, canary promotion and rollback.
Fine-tune jobs
Licence gate and promotion only when beating the prompt-only baseline.
Labelling quality
Consensus QA, inter-annotator agreement and adjudication.
PII redaction
Before labelling, with reviewer access logging.
Tenant adapter training
Tenant-scoped jobs with the Training Gym.
Data Labeling Service
Vetted experts through Human Expert Cloud (W4).
Erasure propagation
Marks rows and triggers retrain or exclusion.
Compute budgets
Per job, with GPU pools planned.
Licence compliance
For datasets and base models.
Workflow
Label to promoted model
- 1
Register a consented dataset
- 2
Redact PII
- 3
Run a labelling project with QA
- 4
Train a candidate model under budget
- 5
Compare with the prompt-only baseline
- 6
Promote by canary with rollback
AI
AI skills and Digital Employees
Assisted pre-labelling
Suggests labels for human annotators.
Disagreement detection
Routes annotator disagreements to adjudication.
Digital Employees that use this module
They act through the same governed Actions as people; risky steps wait for approval.
Frequently asked questions
Is AI Lab in Year 1?
Only as a light internal version. The module is eligible from W2 and the labelling service is planned for W4.
Will fine-tuned models replace prompts automatically?
No. A fine-tuned model is promoted only if it beats the prompt-only baseline, through canary and rollback.
How is erasure handled?
Erasure marks dataset rows and triggers retraining or exclusion according to policy.
See AI Lab & Data Labeling in your workflow
Tell us about your team and the outcomes you want. We will walk you through the apps, the Digital Employees and the plan for your industry.