Inspect models.
Expose failures.
NirikshaID Bench is a controlled synthetic testbed for identity-document extraction, tamper detection, localization, and robustness under visual degradation.
One original layout. Five structured fields. Held-out digit-edit evaluation.
Privacy constraints should not make evaluation opaque.
Real identity documents are sensitive and difficult to release. Controlled synthetic data makes labels, perturbations, and split logic auditable—but does not erase the synthetic-to-real domain gap.
Extraction
Measures whether visible names, dates, and identifiers survive structured parsing.
Tamper detection
Separately measures whether a model identifies visual inconsistency and its type.
Generalization
Excludes digit edits from tuning splits, then reports their recall independently.
Every transformation is traceable.
Numbers loaded from real local artifacts.
Measured offline baseline
Measured Qwen and OCR runs are loaded from the versioned result artifacts.
Designed to make leakage difficult.
Seeded generation
Every label records the document seed and generator version for replay.
Schema enforcement
Pydantic validates tamper state, identity format, boxes, images, and split membership.
Common contract
Adapters emit the same strict JSON structure with parser and latency metadata.
NirikshaID Bench evaluates extraction and forgery classification alongside paired intervention sensitivity, hard-negative false positives, held-out tamper generalization, and uncertainty calibration.
Useful evidence, bounded claims.
Domain gap
Synthetic cards cannot represent the full distribution of real capture conditions.
Roadmap
Multiple layouts, multilingual text, screen recapture, subtler forgeries, and private validation.
Optional tuning
QLoRA configuration is isolated from the critical evaluation path and never sees digit edits.
Ninad Naik
CS Student | AI/ML Research and Engineering
Interested in large language models, multimodal learning, document intelligence, computer vision, and reproducible AI evaluation.