Video write-up · May 7, 2024
V7 Labs data labeling
A document-labeling experiment that compares model outputs, adds human evaluation and exports the resulting structured data.
The V7 Go walkthrough uses handwritten saffron-farming notes as a realistic messy input. Images are dropped into a project, OCR creates an initial text layer, and generative models are asked to convert the material into a more logical structure. The interface can place outputs from different models beside the same source rather than hiding the comparison in separate runs.
The most useful feature is the evaluation loop. Pratik compares GPT-4 and Claude outputs, selects or recomputes results, and adds a consistency check. The project can then export a CSV containing the original file reference, OCR text, model outputs and the human judgment. That turns labeling into an auditable dataset-building process instead of a one-shot model answer.
The demo points toward a practical pattern for document operations: ingest the source, preserve the raw extraction, generate competing structured interpretations, record human preference, and only then export. Webhooks can connect the approved result to tools such as Zapier, but the approval and provenance remain visible before automation continues.
What to take from it
- Keep source files, OCR and generated labels together for traceability.
- Compare models on the same inputs instead of trusting a single output.
- Export human judgments with the data so evaluation becomes reusable training evidence.