Ground truth is the verified, human-confirmed reference dataset that an AI extraction system is measured against to decide whether its output is correct. In commercial real estate document extraction, ground truth is the set of known-correct field values from a lease, offering memorandum, or rent roll, established by expert review before any accuracy metric can be computed.
What Is Ground Truth in Document Extraction?
Ground truth is the set of correct answers held aside to grade a model's output. In CRE extraction it is the base rents, commencement dates, square footage, and clause terms that a human reviewer has confirmed against the source document. Without it, no accuracy, precision, or recall figure can be calculated, because there is nothing to compare against.
Ground truth is produced by annotation: a person reads the source document and records the true value for each field. Labeled data becomes ground truth only when it is accurate, validated, consistent, and reliable enough to serve as a reference standard (Label Your Data). A single annotator can introduce error or bias, so higher-stakes datasets use multiple reviewers and keep a value only when they agree.
Component | What it holds |
Source document | The original lease, OM, or rent roll page |
Field labels | The list of values to be captured, such as base rent or lease term |
Verified values | The human-confirmed correct answer for each field |
Location data | Where on the page each value sits, for span-level checking |
Why Ground Truth Matters
Ground truth matters because the quality of every accuracy claim is capped by the quality of the reference behind it. A model reported as 98% accurate is only 98% accurate against a specific ground truth set. If that set is wrong or inconsistent, the metric is meaningless, and errors in the reference propagate directly into training and evaluation.
The reliability gate is inter-annotator agreement, the rate at which independent reviewers assign the same label. Human labels are often used as ground truth in AI benchmarks, but when human labels diverge, model training becomes unstable and benchmarks weaken (iMerit). In CRE this is concrete: a co-tenancy trigger or a gross-up provision can be read two ways by two analysts, and if the ground truth itself is contested, the score built on it inherits that ambiguity. Ground truth is the reference every other number depends on, which is why it is built once, carefully, and treated as fixed.
Example
Ground truth is easiest to see when building a test set for a lease-extraction model. A team selects 20 retail leases and asks two senior analysts to independently record 40 fields per lease, for 800 field values total. The values are compared, and only agreed values become ground truth.
Step | Count | Note |
Fields labeled per lease | 40 | Rents, dates, square footage, clause terms |
Leases in the set | 20 | Representative mix of clean and amended |
Total field values | 800 | 40 x 20 |
Values where both analysts agreed | 772 | Accepted directly as ground truth |
Values where they disagreed | 28 | Adjudicated by a third reviewer |
Inter-annotator agreement here is 772 / 800 = 96.5%. The 28 disagreements, mostly buried renewal terms and narrative clauses, are resolved by a third analyst before the set is frozen. The finished 800-value reference is now the fixed benchmark: any model run against these 20 leases is scored on how closely it reproduces these confirmed values, not on anyone's opinion after the fact.
Variations and Edge Cases
Ground truth is not a single thing. Its form changes with the task and its trustworthiness changes with how it was built. The variants below alter what a given reference set can be relied on to prove.
Variant | Behavior |
Gold standard set | Small, expert-reviewed benchmark used for final evaluation |
Silver standard set | Larger, machine-assisted labels; cheaper but less certain |
Single-annotator | Fast but carries one reviewer's errors and bias |
Multi-annotator with adjudication | Slower; disagreements resolved, higher trust |
Ambiguous-field ground truth | Fields with no single correct reading; agreement is inherently low |
Ground Truth vs Training Data
Ground truth is often confused with training data, but they play different roles. Training data is the labeled examples a model learns patterns from. Ground truth is the verified reference a model is graded against. The same labeled records can serve both purposes, but never in the same split: data used to train a model cannot honestly measure it.
Put simply, training data teaches and ground truth tests. In a sound evaluation the two are separated so the test set stays unseen during training. A model scored against leases it was trained on reports inflated accuracy, because it is being asked to recall answers it has already memorized rather than to read documents it has never seen.
Frequently Asked Questions
What is ground truth in document extraction?Ground truth is the set of verified, human-confirmed correct values for a document, held aside as the reference an AI system is measured against. In CRE it is the confirmed rents, dates, and clause terms from a lease or rent roll. No accuracy figure can exist without it.
How is ground truth created?Ground truth is created by human annotation: reviewers read the source document and record the true value for each field. Higher-trust sets use multiple independent annotators and keep a value only when they agree, using inter-annotator agreement as a quality gate.
What is the difference between ground truth and training data?Training data is what a model learns from; ground truth is what it is graded against. They must be kept in separate splits, because a model tested on the same data it trained on reports inflated, untrustworthy accuracy.
Related Terms
Extraction Accuracy
Precision and Recall
Intelligent Document Processing
Human-in-the-Loop
Rent Roll