PDF to Structured JSON — 3x faster than Gemini

Datalab introduces lift, an open weights 9 billion parameter model that takes a pdf and a set of fields to extract and outputs a valid json object. In a test set of 225 documents spanning 6 page invoices to 64 page reports (~11,000 scored fields), it achieves 90.2% field accuracy compared to NuExtract3 (81.5%) and Qwen3.5-9B (76.3%).

The key features of lift that make it applicable to real world scenarios are: (1) Schema constraint: The model is forced to produce schema constrained json objects mechanistically rather than being prompted to do so. The output will always adhere to the schema provided. (2) Trained abstention: If the model does not find the required field in the pdf, it will abstain and output null rather than hallucinate. This is crucial for real world deployment because it is better for the ai model to be certain about a null value than incorrect about a value. For example, it would be better if the ai model was certain about a null value for invoice amount rather than being certain about an incorrect invoice amount. In this case, you may not realize the mistake until after funds have been transferred.

1What this means if you work with documents

Consider all the processes that involve working with pdfs: entering invoices into accounting systems, submitting insurance claims, indexing research articles based on date and author, analyzing bank statements to extract transaction information. All of these things are typically done manually or through fragile scripts that fall apart whenever the pdf layout is changed. With lift, you can automate these processes. Simply tell the model what fields you want it to extract (e.g., invoice_number, vendor_name, total_amount) and it will provide those fields for any pdf you feed it (with a median response time of 9.5 seconds per document).

lift outperforms cloud apis such as Gemini Flash 3.5 in terms of speed (3x) for performing similar pdf extraction tasks. It also runs locally on your device. The source code is licensed under apache 2.0 while the model weights are licensed under modified openrail-m. There are no charges per document and no data is sent outside of your server, ensuring privacy of sensitive financial/customer details contained within the pdf files.