datalab-to/lift · Hugging Face

Source: https://huggingface.co/datalab-to/lift

lift is a structured extraction model fromDatalabthat pulls structured JSON out of PDFs and images. Pass any JSON schema and lift returns a JSON object matching it, using schema-constrained decoding to guarantee valid, well-typed output.

Try lift in thefree playground, or use thehosted APIfor higher accuracy, per-field verification, and citations.

https://huggingface.co/datalab-to/lift#featuresFeatures

• Extract structured data from documents
• Pass any JSON schema
• Handles multi-page documents in a single pass, including values that span pages
• Two inference modes: local (HuggingFace) and remote (vLLM server)
• CLI for single files, inline schemas, or whole directories
• Schema Studio: a Streamlit app to build, save, and test schemas against your documents

https://huggingface.co/datalab-to/lift#quickstartQuickstart

pip install lift-pdf

# With vLLM (recommended, lightweight install)
lift_vllm
lift_extract input.pdf ./output --schema schema.json

# With HuggingFace (requires torch)
pip install lift-pdf[hf]
lift_extract input.pdf ./output --schema schema.json --method hf

A schema is standard JSON Schema. Keep it simple —string,number,integer,boolean, arrays of those, arrays of objects, and nested objects are all supported. Write adescriptionfor any field whose name isn’t self-explanatory, and mark a fieldrequiredonly when it must appear; fields genuinely absent from a document come backnull.

{
  "type": "object",
  "properties": {
    "invoice_number": {"type": "string", "description": "Invoice identifier"},
    "total": {"type": "number", "description": "Total amount due"},
    "line_items": {
      "type": "array",
      "items": {
        "type": "object",
        "properties": {
          "description": {"type": "string"},
          "amount": {"type": "number"}
        }
      }
    }
  },
  "required": ["invoice_number", "total"]
}

https://huggingface.co/datalab-to/lift#usageUsage

https://huggingface.co/datalab-to/lift#with-vllm-recommendedWith vLLM (recommended)

from lift import extract
from lift.model import InferenceManager

# Start the vLLM server first with: lift_vllm
model = InferenceManager(method="vllm")
result = extract("document.pdf", "schema.json", model=model)
print(result.extraction)

https://huggingface.co/datalab-to/lift#with-huggingface-transformersWith HuggingFace Transformers

from lift import extract
from lift.model import InferenceManager

# Loads datalab-to/lift in-process (requires: pip install lift-pdf[hf])
model = InferenceManager(method="hf")
result = extract("document.pdf", "schema.json", model=model)
print(result.extraction)

extractaccepts the schema as a dict, a path to a\.jsonfile, an inline JSON string, or the name of a saved schema. Passpage\_range="0\-5"to limit PDF pages, and setVLLM\_API\_BASEto target a remote server.

https://huggingface.co/datalab-to/lift#benchmarksBenchmarks

Evaluated on a 225-document extraction benchmark (6–64 pages per document, ~11,000 scored fields) with adversarial cases planted throughout: cross-page values, exhaustive lists, fields that must be left null, near-miss distractors, multi-source aggregation. Scoring is deterministic exact-match against ground truth (numeric tolerance, normalized strings).

All models receive the same rendered page images, and extract each document in a single pass.

ModelSizeField accuracyFull-document accuracyMedian latency*FeaturesDatalab API—95.9%44.4%30.8sCitations + VerificationGemini Flash 3.5—91.3%40.0%28.1slift9B**90.2%**20.9%9.5sAzure Content Understanding—83.4%22.2%73.7sNuExtract34B81.5%8.4%8.3sQwen3.5-9B9B76.3%24.0%16.8s * Per document, 8 concurrent requests. Local models (lift, Qwen3.5-9B, NuExtract3) served with vLLM on a single GPU; Gemini, Datalab, and Azure via API. Latency varies with hardware and load — treat as relative, not absolute.

• Field accuracy— fraction of individual schema fields extracted correctly.
• Full-document accuracy— fraction of documents whereeveryfield is correct.

Hosted models with verification, citations, and confidence scores are available via theDatalab API— test in theplayground.

https://huggingface.co/datalab-to/lift#commercial-usageCommercial Usage

Code is Apache 2.0. Model weights use a modified OpenRAIL-M license: free for research, personal use, and startups under $5M funding/revenue. Cannot be used competitively with our API. For broader commercial licensing, seepricing.

https://huggingface.co/datalab-to/lift#creditsCredits

• Huggingface Transformers
• vLLM
• Qwen 3.5