Datalab releases lift, an open-weights 9B model that extracts JSON from PDFs

Datalab has released lift, an open 9-billion-parameter vision model that converts PDF documents and images into structured JSON following a given schema. On a set of 225 real-world documents, the model reached 90.2% field accuracy, one of the key metrics in automated document parsing.

How lift works

The principle is straightforward: the model takes two inputs, a document (PDF or image) and a JSON schema describing the desired data structure. The output is a JSON object whose fields are filled with values extracted from the document.

The key technical feature is schema-constrained decoding. At each token-generation step, the allowed options are limited to those that match the given schema. The resulting JSON is always syntactically valid and matches the expected types: no stray brackets, no string where a number is expected, no array where an object is expected.

The second important feature is trained abstention. If a field from the schema is absent from the document, the model explicitly returns null instead of inventing a value. This directly addresses hallucinations: generative models often "fill in" empty fields with plausible but fabricated data. An explicit null is more reliable, since a downstream system can handle a missing field correctly instead of receiving a silent error.

What the benchmark showed

Datalab tested lift on a set of 225 real-world documents of various types. The main metric is field accuracy: the share of fields the model filled in correctly against the reference annotation. The final result is 90.2%, a solid figure for document intelligence.

Context matters here: a single document type (an invoice, for example) can exist in dozens of formats from different vendors, with differently structured tables, handwritten notes and poor scan quality. High accuracy on a heterogeneous set signals real generalization.

Key characteristics of the model:

• Open weights: deploy without depending on a third-party cloud API
• 9 billion parameters: fits on a single server GPU (A100, H100) or a powerful consumer card
• Arbitrary JSON schemas: adapts to any document type without fine-tuning
• Returns null instead of hallucinating: predictable behavior when the source lacks the data
• Native PDF and image handling, with no separate OCR stage at the input

Why businesses need it

Parsing unstructured documents is a chronic pain point in corporate processes. Supplier invoices, contracts, medical records, customs declarations, bank statements and insurance policies all arrive in different formats and require either manual data entry or expensive automation.

The traditional approach relies on a multi-stage pipeline: OCR for text recognition, normalization, named-entity extraction via NLP, post-processing and manual verification of doubtful fields. Each stage is a separate point of failure and a separate line item in development and maintenance costs.

lift shortens that path: you describe the structure you need as a JSON schema, pass in the document and get finished JSON back. Open weights let you deploy the model in your own environment and, if needed, fine-tune it on corporate documents without sending data to external services. For financial institutions, medical facilities and law firms with strict confidentiality requirements, this is essential.

Datalab is already known in the community for Marker, a high-quality PDF-to-Markdown converter. lift continues that line, adding structured output and strict data typing.

What it means

Specialized open models for document intelligence lower the barrier to automating document workflows. If lift holds its claimed accuracy on real corporate data, it becomes a serious alternative to cloud platforms such as Amazon Textract or Azure Form Recognizer, without vendor lock-in and with the option to fine-tune for your own document types.