Attn_exp_data

Frozen public-data inputs for the GroundGuard attention experiments, seed 42. This is a curated experimental subset of VistaQA, GSEval, GroundingME, and HalluSegBench. All selected units are used for testing. There is no separate training, development, or pilot split. Only the selected records and referenced image/mask assets are packaged. Images retain their original resolution.

Dataset	Test units	Image/query records
VistaQA	1,157	1,157
GSEval	1,000	1,000
GroundingME	855	855
HalluSegBench	650 pairs	2,600
Total	3,662	5,612

HalluSegBench contains all four conditions per pair: factual_original, factual_replacement, counterfactual_original, and counterfactual_replacement. There are 325 reasoning pairs and 325 referring pairs; no extra validation pairs are included. The package has 5,559 image/PNG-mask files, approximately 2.51 GB before Git metadata.

Small pilot runs use a runtime sample limit. These samples remain in the full analysis; there are no pilot labels or additional pilot manifests. Run each experiment on a small number of samples from each applicable dataset first, then remove the runtime limit for the full analysis.

Download and use

Download inside the experiment project's root directory:

python -m pip install -U huggingface_hub
hf download SteveWCG/Attn_exp_data --repo-type dataset --local-dir Attn_exp_data

Point the existing experiments at the curated data:

python -m experiments --data-root Attn_exp_data/prepared_data prepare
python -m experiments --data-root Attn_exp_data/prepared_data plan --experiment 1 --subset test

Use the same --data-root for subsequent commands. The package's manifests and build report have new hashes because unused records were removed. Existing work directories created from a full candidate catalog must be regenerated in a fresh work directory. The selected unit IDs and retained records are unchanged; every selected unit is now in test. Regenerate work directories created under the old 90%/10% split.

The included lightweight reader also works independently with Python 3.11 or later:

cd Attn_exp_data
python -m pip install -r requirements.txt
python scripts/verify_data.py

from groundguard_data import Dataset

data = Dataset('prepared_data')
for sample in data.iter_samples('vistaqa', subset='test'):
    inputs = data.model_input(sample)  # image + question only
    reference = data.reference(sample)
    mask = data.region_mask(sample)

sample = next(data.iter_samples('hallusegbench', subset='test'))
conditions = data.get_pair(sample)  # all four image/query combinations

Layout and data contract

prepared_data/
  manifests/<dataset>.jsonl   Retained normalized records
  selections/<dataset>.json   Frozen selected IDs; test includes every selected ID
  images/<dataset>/          Referenced images only
  masks/<dataset>/           Referenced binary PNG masks only
  build_report.json          Counts, source revisions, manifest checksums
groundguard_data/            Reader; no model dependencies
scripts/verify_data.py       File/hash, image, mask, split and pair checks
scripts/export_experiment_data.py  Recreate export from full prepared catalog
data_config.json             Current preparation/selection policy
sources/                    Dataset cards from pinned upstream snapshots
SHA256SUMS                  SHA-256 of every other repository payload file
verification_report.json    Results from validating this export

sample_id identifies an image/query condition; unit_id identifies a native item or HalluSegBench pair. source_group_id is the base image's decoded RGB hash. All paths are relative to prepared_data/. reference holds evaluation labels; provenance retains upstream repository revision, annotation location, native annotation, and any recorded coordinate normalization. Native annotation paths are historical references to the upstream archives, not required local assets. Coordinates are original-image bbox_xyxy pixels; mask sizes are [height, width]. COCO RLE masks remain inline in the JSONL; PNG masks are separate assets. intervention_regions retain the edited object's geometry in counterfactuals.

subset='selected' and subset='test' return the same 1,157 / 1,000 / 855 / 650 primary units. development_unit_ids is an empty compatibility field, not a held-out split. subset='all' means the packaged catalog, not the original candidate pool. Primary conditions are the default; use condition='all' to iterate every counterfactual condition.

GSEval retains four classes with 250 distinct source images per class. GroundingME excludes the Small subtask. VistaQA's original duplicate records remain intact. The upstream invalid HalluSegBench pair refer_seg/298954 is not packaged; historical exclusions remain in the selection files for provenance. All variants of a source scene retain their shared grouping for analysis.

No raw archives, unused candidates, model weights, generated experiment outputs, local caches, or credentials are included. This repository contains only the four public datasets. The experiment project's prepare command freezes a 500-scene action audit from these public data; this is an analysis subset, not a development or pilot split. Synthetic checks are deferred until after the public-data results and are not part of this upload or the default preparation step.

Provenance and attribution

Source	Pinned revision	Original card
VistaQA	cf0cb9857e5f0969a7fc2c8efb3ae9d4816ccc1a	Card and citation
GSEval	6e22511a0d873f98c506dd9a16edc335b642615b	Card and citation
GroundingME	78e3c7974b2b1db0ea52266969e40f664a38e330	Card and citation
HalluSegBench	634dd7ecdb7a46a457a2f6091f3185dcbf1e5bad	Card and citation

This curation does not replace upstream licenses or ownership. See each preserved source card and upstream dataset for its applicable terms and citations. VistaQA declares CC BY 4.0; GSEval declares Apache 2.0; GroundingME declares other with details in its card. HalluSegBench's pinned card does not declare a license.