# AI Training Data From Real Workflows

> MARFI Trajectory converts consented workplace sessions into state-action-state training data for AI agents and workflow models.

Canonical URL: https://insight.marfi.app/ai-training-data

AI teams need demonstrations that reflect real tools, real state, and real user actions. MARFI Trajectory captures consented work sessions and exports aligned state-action-state examples that can feed model training, evaluation, and workflow automation research.

## Audience

AI product teams, automation teams, and applied research groups building task agents.

## Search Intent

Find proprietary AI training data from real software workflows and user demonstrations.

## Every trajectory preserves context.

A useful training example is more than a click. MARFI aligns before state, action, after state, app context, window or URL, and system state so downstream models can learn the conditions surrounding a workflow step.

- State before an action
- Human action or semantic event
- State after the action
- Optional keyframes when screenshot consent is present

## The dataset belongs to the customer.

Raw captures and derived trajectories are customer-owned. MARFI takes no license to customer work data and does not train its own models on it.

## Move from work signal to training pipeline.

Trajectory workspaces can export clean JSONL examples and kick off hosted fine-tuning paths where supported. Other providers receive generic training records that teams can transform into their preferred format.

## FAQ

### What makes MARFI training data different from synthetic examples?

MARFI starts from real work sessions on real software, with state and action context captured from enrolled machines under explicit consent controls.

### Can MARFI train on my customer data?

No. MARFI does not train its own models on customer data and takes no license to raw captures or derived trajectories.

## Related Pages

- [State-Action-State Trajectory Data](https://insight.marfi.app/state-action-trajectory-data): State-action-state trajectory data from MARFI connects before state, human action, and after state for AI agent training workflows.
- [Consent-Gated Screenshot Capture](https://insight.marfi.app/consent-gated-screenshot-capture): MARFI screenshot capture is Trajectory-only, consent-gated, and designed for visual replay and AI training trajectories.
- [Data Ownership](https://insight.marfi.app/data-ownership): MARFI customers own their raw work captures and every derived trajectory. MARFI takes no license and does not train models on customer data.
- [AI Teams](https://insight.marfi.app/solutions/ai-teams): MARFI Trajectory helps AI teams collect consented workflow demonstrations and export state-action-state JSONL for agent training.