Datalab's lift: The 9B Model That Knows When to Shut Up

This open-weights extractor isn't just accurate, it's honest about what it doesn't know

By AI·Reporter·June 23, 2026·~4 min read

Takeaways

›lift uses schema-constrained decoding to guarantee valid JSON structure, not correctness
›Its key innovation is trained abstention: returning null for uncertain fields
›Strong on field-level accuracy, weaker on full-document perfection
›Best for human-in-the-loop review and analytics, not fully automated processing

Document AI is crowded with models that hallucinate. Datalab's new lift model takes a different tack: it shuts up when it's not sure. This 9B-parameter vision model extracts structured data from PDFs and images, but its real innovation is knowing when to return null.

The Core Trick: Decode Against the Schema

lift's key mechanism is schema-constrained decoding. You give it a JSON schema, and it generates directly against that structure, token by token. Here's the flow:

Your schema becomes a strict JSON Schema
That schema constrains the vLLM inference server
During generation, only schema-valid tokens are allowed

This guarantees structural validity, not correctness. A 'number' field will always be a number, but it might be the wrong number.

json

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

This invoice schema will always produce valid JSON. But validity isn't truth.

Abstention: The Real Superpower

lift's trained to return null for fields it can't confidently extract. This matters more than raw accuracy.

Consider invoice processing. A model that invents a tax ID is worse than useless, it's dangerous. lift leaves that field null, signaling clearly: "Not found."

This abstention is baked into both training and decoding. Every field accepts null, letting the model opt out without breaking structure.

The Numbers: Strong on Fields, Weaker on Full Docs

On Datalab's 225-document benchmark:

Model	Size	Field accuracy	Full-doc accuracy	Median latency
lift	9B	90.2%	20.9%	9.5s

lift leads self-hostable models in field accuracy. It's within spitting distance of Gemini Flash 3.5 (91.3%) while running 3x faster. But its full-document accuracy, where every field must be right, lags at 20.9%.

This isn't just lift's problem. Even the best model (Datalab's own API) only hits 44.4% full-document accuracy. Perfect extraction remains hard.

Where It Fits: Review and Analytics, Not Full Automation

lift shines for field-level extraction feeding human review or analytics. It's less suited for zero-touch automation where every field must be perfect.

In contract review, it can stitch values across pages. In accounts payable, null due dates prevent silent errors. But for fully automated, high-stakes extraction? You'll need more than lift alone.

The Honest Extractor

lift's innovation isn't perfect extraction. It's knowing when to say "I don't know", a crucial skill for turning document chaos into reliable data. In a field full of overconfident models, lift's restraint is its superpower.

This 9B-parameter model, with Apache 2.0 code and modified OpenRAIL-M weights, pushes open-source document AI forward. Not by being omniscient, but by being honest about its limits.