SQuID-Train — 75,687 Quantitative Spatial Reasoning Training Pairs

Training companion to the SQuID benchmark (https://huggingface.co/datasets/squid-bench-anon/SQuID), under review at the NeurIPS 2026 Evaluations and Datasets Track.

• 75,687 question-answer pairs over 1,443 satellite images
• Generated by the same pipeline as the benchmark (identical question templates, geometric definitions, minimum-area thresholds, and GSD handling), applied to the training partitions of the source datasets (DeepGlobe, EarthVQA, PV) — zero image overlap with the 2,000-question evaluation benchmark, which is built exclusively from published validation/test partitions.
• Schema: image, question, answer, type, tier, dataset, id (exact targets; no tolerance ranges — those are an evaluation-side concept).
• Tier distribution: Tier 1 = 19,845, Tier 2 = 50,509, Tier 3 = 5,333; 7,146 zero-valued answers for feature-absence robustness.
• Files: qvlm_benchmark_training.json, dataset_statistics.json, images_and_masks/ (images + masks for the training partition).

Reference result

LoRA-finetuning Qwen2.5-VL-7B-Instruct on this split raises its SQuID benchmark accuracy from 16.1% to 40.7% (Tier 1: 53.5%, Tier 2: 56.0%, Tier 3: 13.2%) — end-to-end models can learn substantial metric perception from this data, while multi-step compositional reasoning (Tier 3) remains open.

Research use only (CC BY-NC 4.0; imagery licenses inherited from the public source datasets).