Datalab Launches lift: A 9B Parameter Vision Model for JSON Extraction from PDFs

Datalab has unveiled lift, a vision model with 9 billion parameters focused on structured data extraction. By utilizing a JSON schema as input, lift processes PDFs and images to generate a corresponding JSON object. This model marks Datalab's initial foray into dedicated extraction technology, building on their existing open-source OCR tools like chandra, marker, and surya. With a field accuracy of 90.2% on a benchmark of 225 documents, lift is touted as the most effective small self-hostable model available, processing each document in an average of 9.5 seconds. The model supports multi-page documents in a single operation, allowing it to read values across pages seamlessly. It offers two inference modes: local inference through HuggingFace and remote inference via a vLLM server, the latter being recommended for production environments. The model's architecture incorporates schema-constrained decoding, ensuring that the output adheres to the specified JSON structure. However, users should be cautious of certain schema constructs that may not compile correctly, as this can lead to unstructured output. Overall, lift is positioned as a valuable tool for extracting structured data while minimizing the risk of erroneous field generation.