SQuID-GT — Ground-Truth Mask Bundle for the SQuID Benchmark

Companion release for SQuID (https://huggingface.co/datasets/squid-bench-anon/SQuID), under review at the NeurIPS 2026 Evaluations and Datasets Track. The evaluation split on the main repository is unchanged; this bundle adds the assets needed for diagnostic (oracle) evaluation with the QASR code release (https://anonymous.4open.science/r/qasr-squid/):

• images_and_masks/{deepglobe_0.5m, earthvqa_0.3m, solar_0.3m, naip_1.0m}/ — source images and the curated segmentation masks that SQuID ground-truth answers are derived from (1,950 mask-backed questions; the 50 NAIP questions use human-consensus ground truth and have no masks).
• qvlm_benchmark.json — the benchmark records with per-question image/mask paths, answers, tolerance ranges, tiers, and GSD (refreshed export with unified pixel-count area definitions and content-hash ids).

Purpose

Running QASR with these ground-truth masks in place of any segmentation model (USE_GT_MASKS=1 in the code release) measures the reasoning + geometry-API ceiling under perfect perception, decomposing benchmark error into perception vs. reasoning. This oracle configuration is fully reproducible from public assets (no closed-source segmentation checkpoints required).

# with the code release checked out:
mv <this download> ./SQuID_GT
./run_oracle_gt.sh validate     # no API calls; expect 1950/1950 and 100% pass rates
OPENAI_API_KEY=sk-... ./run_oracle_gt.sh run

Imagery/masks originate from DeepGlobe (CVPR-W 2018), EarthVQA (ICCV 2023), and a public photovoltaic-panel segmentation dataset (ESSD 2021); NAIP imagery is USGS public domain. Research use only (CC BY-NC 4.0).