GNWT-Bench (Open 50-Question Subset)

This repository hosts an open subset of GNWT-Bench, a perturbation-robustness benchmark. The 50 questions here were drawn from the full benchmark by stratified sampling that preserves the original category proportions. Each question ships with its clean input and 3 perturbation types × 4 severity levels (20/40/60/80) — 13 input conditions per question, 650 inputs in total, plus one metadata JSON per question.

Directory structure

{text|image}/
  {question_type}/
    {question_id}/
      {question_id}.json              # metadata: question, ground truth, provenance, relative paths
      original/0/{id}_0.{txt|png|jpg} # clean input (level 0)
      {perturbation}/{20|40|60|80}/   # perturbed input at that severity

• Text perturbations: antonym, delete, typos
• Image perturbations: contrast, impulse_noise, shot_noise

All paths inside the metadata JSON (paths for text, original_sample / perturbed_samples for image) are relative to the question's folder and expanded per level, e.g.:

"perturbations": {"typos": {"20": "typos/20/text_2364_typos_20.txt", "...": "..."}}

Categories

Modality	Question type	Source dataset(s)	Count
text	objective_choice_discrimination	RACE	10
text	localization_evidence	HotpotQA	8
text	objective_fill_recognition	HotpotQA	7
text	subjective_generation_integration	CNN/DailyMail	7
text	judgment_detection	BoolQ, MultiRC, HotpotQA	7
image	objective_fill_recognition	TextVQA, ST-VQA, VQAv2, COCO-panoptic	7
image	judgment_detection	COCO, VQAv2, COCO-panoptic, TextVQA	2
image	objective_choice_discrimination	CIFAR-10-C	1
image	subjective_generation_integration	COCO Captions	1

Metadata schema

Field	Present	Description
id	all	Question ID (text_NNN / image_NNN)
modality	all	text or image
question_type	all	One of the nine categories above
question	all	Question prompt
gt	all	Ground truth (see note below)
source / source_dataset_info	text / image	Dataset provenance: name, split, source ID
word_count	text	Word count of the clean passage
image_size	image	{width, height} of the clean image
paths	text	Relative paths to clean + perturbed inputs
original_sample, perturbed_samples	image	Relative paths to clean + perturbed inputs

Note on gt: the structure follows the source task — a dict with answer / scoring_method (and answer_label / answer_text for multiple choice) for text questions; a dict, plain string, or list of reference captions for image questions.