Datalab’s New Lift Model Extracts Structured JSON from PDFs with 9B Parameters

Datalab has introduced “lift,” a new 9-billion parameter open-weights vision model engineered to extract structured JSON data from PDFs and images. This model accepts a JSON schema as input and returns a compliant JSON object, directly processing multi-page documents in a single pass. The release marks a significant step in automated data extraction, offering a tool that can be self-hosted and integrated into various data processing workflows.

Key facts

Feature	Detail
Model Name	lift
Parameters	9 Billion
Functionality	Structured JSON extraction from PDFs/images
Key Technology	Schema-constrained decoding, trained abstention
Accuracy (Field)	2% on 225-document benchmark
License	Apache 2.0 (code), OpenRAIL-M (weights)

Enhancing Document Data Extraction

The “lift” model builds upon Datalab’s existing open-source optical character recognition (OCR) tools, such as chandra, marker, and surya. Its core innovation lies in schema-driven field extraction, where it directly decodes output against a user-defined JSON schema. This approach ensures that the resulting JSON is always structurally valid according to the provided schema, minimizing errors often associated with unstructured data extraction. The model is designed to handle complex documents, including values that span multiple pages.

Technical Implementation and Performance

“lift” operates by converting JSON Schemas into Pydantic models, which are then normalized into strict JSON Schemas. This processed schema is used as a constraint during the generation process, where the model assigns probabilities to possible next tokens while masking out any tokens that would violate the schema. This method guarantees structural integrity but highlights that “validity is not correctness”—meaning, while the output will be structurally sound, its factual accuracy still depends on the model’s interpretation of the document content.

The model also features “trained abstention,” allowing it to return null for genuinely missing fields instead of hallucinating data. This functionality is crucial for preventing silent errors in downstream applications, as a model that invents absent data can be more problematic than one that acknowledges missing information. Fields are only marked as required if they must appear, otherwise, the model will abstain if a field is not present in the document.

Benchmark Results and Market Position

On Datalab’s internal 225-document benchmark, “lift” achieved a 90.2% field accuracy. This benchmark involved documents ranging from 6 to 64 pages, with approximately 11,000 scored fields, including adversarial cases designed to test cross-page values and fields requiring null abstention. The model demonstrated strong performance among self-hostable options, outperforming NuExtract3 and Qwen3.5-9B in field accuracy. With a median processing time of 9.5 seconds per document, it also proved to be faster than some leading hosted API solutions like Gemini Flash 3.5, while maintaining comparable accuracy.

While “lift” leads in field accuracy for self-hostable models, its full-document accuracy—where every field must be correct—was 20.9%. This reflects the inherent difficulty of single-pass extraction on lengthy and complex documents, a challenge faced by all models in the benchmark. This shows that despite advances, challenges remain for perfect end-to-end document extraction.

Implications for Indian Enterprises

For businesses and developers in India, “lift” presents an opportunity to automate the extraction of critical structured data from a wide array of documents, including invoices, reports, and legal texts. Its open-weights nature and Apache 2.0 license for code provide flexibility for custom implementations and integration into existing systems without proprietary lock-in. The ability to self-host the model also addresses data privacy and security concerns that are often paramount for Indian enterprises handling sensitive information. This development can accelerate digital transformation efforts and improve efficiency in sectors reliant on paper-based or scanned document workflows.

Source: MarkTechPost: https://www.marktechpost.com/2026/06/23/datalab-releases-lift-a-9b-open-weights-vision-model-that-extracts-structured-json-from-pdfs-using-schemas/