TaughtHuman decisions for better models

Custom model training with a reason to trust the winner.

Connect eligible human-reviewed evidence to reproducible dataset versions, candidate training runs, frozen evaluation, and an explicit promotion decision.

Request early access Browse model releases
Evidence registerTraining completion is not a promotion gate.

Taught keeps the candidate, incumbent, dataset, split, metrics, artifacts, and decision note together.

  1. Freeze
    Version the dataset

    Eligible labels and task rules become a reproducible training input.

  2. Split
    Protect evaluation

    Source groups help prevent near-duplicate leakage across train and test.

  3. Train
    Record the run

    Architecture, parameters, artifacts, and evidence remain attributable.

  4. Promote
    Compare and decide

    A frozen gate determines whether the candidate may replace the incumbent.

Turn reviewed data into a controlled model change.

Custom training is most useful when the team can reproduce the data, compare the candidate with the current model, and explain the promotion decision. Taught keeps those objects in one governed workflow.

Dataset manifest
The frozen input identifies eligible evidence, taxonomy, task revision, and split membership.
Leakage-aware groups
Related source frames can remain together so evaluation does not reward memorizing near-duplicates.
Run artifacts
Candidate configuration, output checkpoints, metrics, and logs remain attached to the run.
Promotion record
The incumbent changes only through an explicit decision against frozen gates.

The path from evidence to promoted version

  1. 01
    Freeze eligible labels

    Build a dataset from review evidence that has passed the project’s rules.

  2. 02
    Create protected splits

    Use source-aware groups and fixed manifests to make later comparisons meaningful.

  3. 03
    Train a candidate

    Run a supported native or external training path without changing the frozen test evidence.

  4. 04
    Evaluate before promotion

    Compare the candidate with the incumbent and record the decision that changes the active model.

Training support is task-specific

  • Image object detection has a mature native training and evaluation path.
  • Other labeling tasks may use external training or have no training capability; support is reported per task.
  • A trained checkpoint is not automatically serving-compatible or publicly deployable.

Questions teams ask before they begin

Direct answers based on the current Taught product and deployment boundaries.

Does Taught retrain a model after every label?

No. Eligible evidence is frozen into a dataset version, and training runs are explicit, reproducible operations.

How does Taught reduce train-test leakage?

Acquisition or source groups can keep related frames together when fixed dataset splits are created.

Is the newest model automatically promoted?

No. Promotion requires a recorded comparison against frozen evaluation rules and an explicit decision.

Bring the decisions, data, and model version into one record.

Request access for a labeling project, a custom-model workflow, or an integration with a promoted model.

Request early access