Datalab выпустила lift — 9B-модель с открытыми весами для извлечения JSON из PDF

Datalab released lift — an open visual model with 9 billion parameters that transforms PDF documents and images into structured JSON according to a given schema. On a dataset of 225 real documents, the model achieved 90.2% field accuracy — one of the key metrics in automatic document parsing tasks.

How lift Works

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

The key technical feature is schema-constrained decoding. At each step of token generation, allowed variants are restricted only to those that comply with the given schema. The resulting JSON is always syntactically valid and matches the expected types: no stray brackets, strings instead of numbers, or arrays where objects are expected.

The second important feature is trained abstention. If a field from the schema is absent in the document, the model explicitly returns null instead of inventing a value. This directly solves the hallucination problem: generative models often "fill in" empty fields with plausible but fictional data. An explicit null is more reliable — the downstream system can properly handle the absence of a field rather than receive a silent error.

What the Benchmark Showed

Datalab tested lift on a dataset of 225 real documents of various types. The primary metric is field accuracy: the proportion of fields that the model filled correctly relative to ground truth annotations.

The final result — 90.2% — is a significant indicator for document intelligence. It's important to understand the context: the same document type (for example, an invoice) can exist in dozens of formats from different suppliers, have different table layouts, handwritten notes, and poor scan quality. High accuracy on a heterogeneous dataset is a sign of real generalization capability.

Key model characteristics:

• Open weights — deployment without dependency on third-party cloud APIs
• 9 billion parameters — fits on a single server GPU (A100, H100) or powerful consumer card
• Arbitrary JSON schemas — adapts to any document type without fine-tuning
• Returns null instead of hallucinations — predictable behavior when data is absent from the source
• Native PDF and image processing without a separate OCR stage at the input

Why This Matters for Business

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

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

lift shortens this path: you describe the needed structure as a JSON schema, pass the document — get ready JSON. Open weights allow you to deploy the model in your own infrastructure and, if necessary, fine-tune it on corporate documents without sending data to external services. For financial organizations, medical institutions, and law firms with strict confidentiality requirements, this is fundamentally important.

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

What This Means

Specialized open models for document intelligence lower the barrier to entry for document automation. If lift maintains its claimed accuracy on real corporate data, it becomes a serious alternative to cloud platforms like Amazon Textract or Azure Form Recognizer — without vendor lock-in and with the ability to fine-tune for your own document types.