Datalab’s 9B Vision Model ‘Lift’ Turns PDFs Into JSON—No More Hallucinations

Data extraction from PDFs has long been a bottleneck for businesses, especially when documents come in varied layouts and formats. Datalab’s new **lift** model is poised to change that narrative by converting PDFs and images straight into **structured JSON** that matches predefined schemas.

What is lift?

Released today, **lift** is a **9B**‑parameter, **open‑weights** vision model engineered for document understanding. Unlike traditional OCR pipelines that output plain text, lift delivers JSON objects that conform to user‑supplied schemas, ensuring that every field is validated before it’s returned.

How It Works

At its core, lift blends three innovations:

• Schema‑constrained decoding: The model only generates tokens that fit the target schema, eliminating out‑of‑scope fields.
• Trained abstention: When a field is absent, lift returns null instead of hallucinating data.
• Fine‑tuned on a diverse set of PDFs, images, and form layouts to generalize across industries.

Key Features

• **9B** parameters—balancing performance and inference speed.
• **Open‑weights**: Researchers and developers can access the full model for fine‑tuning.
• **90.2% field accuracy** on a 225‑document benchmark.
• Instant schema validation, reducing downstream data cleaning costs.
• Supports multi‑page PDFs and complex layouts with minimal configuration.

Real‑World Impact

For finance firms, lift can auto‑populate loan agreements and regulatory filings with near‑zero error rates. Legal teams can extract case documents, contracts, and court filings into searchable JSON, cutting manual review time from hours to minutes.

In healthcare, the model can pull patient charts, lab reports, and billing statements into structured records, streamlining compliance with HIPAA and enhancing data interoperability.

Industry Reaction

Tech analysts note that lift’s combination of **schema‑constrained decoding** and **trained abstention** addresses two long‑standing pain points: hallucination and schema drift. Early adopters in the US, UK, and Canada report a 70% reduction in data validation effort.

Whether you’re a data engineer, a product manager, or a compliance officer, **lift** offers a ready‑made solution that turns unstructured PDFs into machine‑ready data—without the usual guesswork. Stay ahead of the curve by integrating this open‑weights model into your data pipeline today.

Ready to transform your document workflows? Explore lift now and discover how it can elevate your data accuracy and efficiency.