lift is a 9B-parameter vision-language model (VLM) specializing in document information extraction, released by Datalab on June 18, 2026. When a user defines the desired output structure as a JSON Schema, it finds the corresponding fields in PDFs and images and returns them as structured JSON. While typical document OCR reads pages as a flat plane of characters, lift is closer to using the JSON Schema as a blueprint and assembling the information that should go into each cell from the entire document. It is designed to handle values that are scattered across multiple pages or span page boundaries within a single processing step, and it offers usage paths through Hugging Face-based local inference, a vLLM server, a command-line interface, and Schema Studio. Existing document processing pipelines often require connecting OCR, layout analysis, field-specific rules, and post-processing validation. Even slight changes in the form require rewriting regular expressions or coordinate rules, and directly using a general-purpose generative model carries the risk of hallucinations (generating plausible but incorrect values) being introduced into the database. The key difference with lift is that it explicitly defines the output contract as a JSON Schema and is a document extraction model trained not to generate missing fields arbitrarily. Therefore, it reduces the intermediate code needed to transform the extracted results to match the key names, and research institutions or companies that require self-hosting can consider deployment configurations that do not send the original documents to external APIs. In life science research, it can be used to normalize data with similar structures but varying formats, such as paper appendices, analysis reports, and experimental logs. For example, a researcher can define sample ID, experimental conditions, measured values, units, and page references in a JSON Schema, process a multi-page PDF, and pass the returned JSON to a pandas or database loading step. The behavior of not generating estimated values when a specific field is missing in the document is particularly important in quality control processes where missing data and actual negative results need to be distinguished. However, the extraction accuracy, the number of pages and image resolutions supported, the schema complexity limitations, and whether field-specific reference coordinates are provided cannot be confirmed solely from the input information, so it should be validated with the official documentation and a representative set of documents before operational deployment. Furthermore, in clinical trial reports or bio-manufacturing quality documents, it can be applied to the initial data extraction step, which combines lot numbers, test methods, acceptance criteria, and results scattered across multiple pages into a single record. Providing the model through a vLLM server allows internal applications to use a common inference endpoint, and Schema Studio can be used as an interface for designing and testing the extraction schema. However, the usage restrictions of the Modified OpenRAIL-M model license, the log and cache policies when processing sensitive documents, and the quantitative performance and hardware requirements require a review of the official documentation. Therefore, the output of lift should be treated as candidate data that goes through original document comparison and field-specific validation rules rather than as automatically confirmed values.

lift is a 9B-parameter vision-language model (VLM) specializing in document information extraction, released by Datalab on June 18, 2026. If the user defines the desired output structure as a JSON Schema, it finds the corresponding fields in PDFs and images and returns them as structured JSON. While typical document OCR reads a page as a flat plane of characters, lift is more akin to having the JSON Schema as a blueprint and finding and assembling the information that should go into each field throughout the document. It can also handle values that are scattered across multiple pages or span page boundaries.

🔬 Standardize Metadata for Experiment Reports: Define sample ID, cell line, treatment concentration, time, measurement value, and units using a JSON Schema to convert multi-page PDFs into JSON and link them with pandas validation rules. Quantitative accuracy and processing speed require public source reconfirmation and internal evaluation.