ImplicitIR

Document images paired with question/answer pairs that require implicit reasoning over the document — deriving, cross-referencing, or computing a value — rather than plain OCR transcription. This is the held-out test split used by the ImplicitIR Benchmark eval harness; the reasoning-annotated training splits (with raw field data and derivation traces, used for SFT/GRPO) live in the ImplicitIR training repo.

Most document images are real-world documents sourced from the Unikie benchmark (MIT licensed); the Synthetic category is procedurally generated and not derived from Unikie. This dataset (questions, answers, and synthetic images) is released under CC-BY-4.0; the images/real/ subset remains additionally subject to Unikie's MIT license and attribution requirement.

Dataset structure

test.jsonl is a doc-level JSONL file: one line per document image, with a list of samples (question/answer pairs) for that document.

{
  "doc_id": "91814768_91814769",
  "image": "images/real/91814768_91814769.png",
  "category": "Administrative",
  "samples": [
    {
      "id": 1,
      "field": "total_to_date_minus_previously_reported",
      "question_type": "derivation",
      "prompt": "This form lists three separate 'Total Expenditures or Disbursements' figures near the bottom: one for This Report, one Previously Reported, and one to Date. By how much does the 'to Date' total exceed the 'Previously Reported' total? Answer with the number only — no currency symbols, commas, or units.",
      "field_present": true,
      "answer": "48085.00"
    }
  ]
}

Fields per sample:

Field	Description
id	Sample index within the document.
field	Short identifier for the field/value the question targets.
question_type	derivation (compute/combine values) or spatial (reason about layout/position/references within the document).
prompt	The question posed about the document.
field_present	Whether the queried field actually exists in the document (always true in this split).
answer	The final answer.

Document-level fields: doc_id, image (path relative to this directory), category (Administrative, Commercial, Advertisement, Accommodation, or Synthetic).

Splits

Split	Documents	Samples
test.jsonl	3,427	4,151

Images live under images/ (3,427 total): images/real/ (427 real-world documents, JPG) and images/synthetic/ (3,000 procedurally generated documents, PNG).

Links

Github:

• ImplicitIR — training code (SFT + GRPO) that consumes this dataset
• ImplicitIR-Benchmark — serves and scores checkpoints trained on this dataset