lift-GGUF

lift is a 9-billion-parameter structured extraction model from Datalab, built on Qwen3.5-9B, that pulls structured JSON data out of PDFs and images by accepting any JSON Schema and returning a matching JSON object via schema-constrained decoding to guarantee valid, well-typed output. It handles multi-page documents in a single pass — including values spanning multiple pages — supports both local (HuggingFace) and remote (vLLM server) inference modes, and ships with a CLI for single files, inline schemas, or whole directories, plus a Streamlit-based Schema Studio for building and testing schemas. On a 225-document benchmark (6–64 pages, ~11,000 scored fields) featuring adversarial cases like cross-page values, exhaustive lists, and near-miss distractors, lift achieves 90.2% field accuracy and 20.9% full-document accuracy at a 9.5s median latency, outperforming similarly-sized open models like Qwen3.5-9B (76.3%) and NuExtract3 (81.5%) while trailing larger hosted systems like the Datalab API (95.9%) and Gemini Flash 3.5 (91.3%); code is Apache 2.0 licensed while model weights use a modified OpenRAIL-M license that's free for research, personal use, and startups under $5M in funding or revenue.

Model Files

llama.cpp