SQuID v2.0: benchmark, training split and ground-truth masks

README.md

@@ -1,232 +1,99 @@
 ---
-license: cc-by-4.0
+license: cc-by-nc-4.0
+pretty_name: SQuID
 task_categories:
 - visual-question-answering
 language:
 - en
 tags:
+- remote-sensing
 - satellite-imagery
 - spatial-reasoning
-- benchmark
 - quantitative-reasoning
+- benchmark
 size_categories:
-- 1K<n<10K
+- 10K<n<100K
+configs:
+- config_name: benchmark
+  default: true
+  data_files:
+  - split: test
+    path: parquet/benchmark/*.parquet
+- config_name: train
+  data_files:
+  - split: train
+    path: parquet/train/*.parquet
 ---
 
-# SQuID: Satellite Quantitative Intelligence Dataset
-
-A comprehensive benchmark for evaluating quantitative spatial reasoning in Vision-Language Models using satellite imagery.
-
-## Dataset Overview
-
-- **2000 questions** testing spatial reasoning on satellite imagery
-- **639 unique images** across four datasets
-- **1950 auto-labeled** questions from segmentation masks (DeepGlobe, EarthVQA, Solar Panels)
-- **50 human-annotated** questions from NAIP imagery with consensus answers
-- **1544 questions** include human-agreement ranges for numeric answers
-- **3 difficulty tiers**: Basic (713), Spatial (610), Complex (677)
-- **3 resolution levels**: 0.3m, 0.5m, 1.0m GSD
+<p align="center">
+  <img src="https://huggingface.co/datasets/PeterAM4/SQuID/resolve/main/assets/logo.png" width="90">
+</p>
 
-## Human Annotation & Agreement Methodology
+<h1 align="center">SQuID: Satellite Quantitative Intelligence Dataset</h1>
 
-### Human Annotation Process
-- **50 questions** on NAIP 1.0m GSD imagery were annotated by humans
-- **10 annotators per question** resulting in 500 total annotations
-- **Answer aggregation**:
-  - **Numeric questions**: Used MEDIAN of all responses for robustness
-  - **Categorical questions** (connected/fragmented): Used MAJORITY voting
-  - **Binary questions**: Converted yes/no to 1/0 and used majority
-
-### Human Agreement Quantification
-From the 500 human annotations, we computed the Mean Median Absolute Deviation (MAD) for each question type:
-- **Percentage questions**: MAD = ±1.74 percentage points
-- **Proximity questions**: MAD = ±2.25 percentage points
-- **Count questions**: Normalized MADc = 0.19 (proportional to count magnitude)
-
-For count questions, we use a normalized MAD (MADc) that makes the acceptable range proportional to the count value:
-```
-MADc = median(|Xi - median(X)|) / median(X) = 0.19
-```
+<p align="center">
+  <a href="https://arxiv.org/abs/2601.13401"><img src="https://img.shields.io/badge/arXiv-2601.13401-b31b1b.svg"></a>
+  <a href="https://github.com/PeterAMassih/qvlm-squid"><img src="https://img.shields.io/badge/GitHub-code-181717.svg?logo=github"></a>
+  <a href="https://doi.org/10.57967/hf/7565"><img src="https://img.shields.io/badge/DOI-10.57967%2Fhf%2F7565-blue.svg"></a>
+</p>
 
-### Acceptable Range Calculation
-These MAD values were applied to ALL numeric questions in the benchmark to define acceptable ranges:
+SQuID tests whether vision-language models can **measure** satellite images: count objects, compute areas in hectares, and reason about distances. Every answer is computed from pixel-level masks, and numeric answers are scored against tolerance ranges calibrated on human annotators.
 
-```python
-import math
-
-# For percentage questions (absolute deviation)
-if question_type == 'percentage':
-    lower = max(0.0, answer - 1.74)
-    upper = min(100.0, answer + 1.74)
-    
-# For count questions (proportional deviation)
-# range(C) = [C - max(1, C × MADc), C + max(1, C × MADc)]
-elif question_type in ['count', 'building_proximity', 'building_flood_risk', 
-                        'building_fire_risk', 'connectivity']:
-    MADc = 0.19
-    dr = max(1, answer * MADc)  # At least ±1 deviation
-    lower = max(0, math.floor(answer - dr))
-    upper = math.ceil(answer + dr)
-    
-# For proximity percentage questions (absolute deviation)
-elif 'within' in question and 'm of' in question:
-    lower = max(0.0, answer - 2.25)
-    upper = min(100.0, answer + 2.25)
-```
+![Examples](https://huggingface.co/datasets/PeterAM4/SQuID/resolve/main/assets/examples.jpg)
 
-**Example count ranges with MADc = 0.19:**
-- C=5 → range [4, 7]
-- C=10 → range [8, 12]
-- C=50 → range [40, 60]
-- C=100 → range [81, 120]
-
-Special cases:
-- Zero values have no range (exact match required)
-- Binary/fragmentation questions have no range (exact match)
-- Ranges are capped at valid bounds (0-100 for percentages, ≥0 for counts)
-
-## Question Types
-
-The benchmark includes 25 distinct question types organized into three tiers:
-
-### Tier 1: Basic Questions (713 questions)
-- **percentage**: Coverage percentage of a land use class
-- **count**: Number of separate regions or objects
-- **size**: Area measurements of regions
-- **total_area**: Total area covered by a class
-- **binary_comparison**: Comparing quantities between two classes
-- **binary_presence**: Checking if a class exists
-- **binary_threshold**: Testing if values exceed thresholds
-- **binary_multiple**: Checking for multiple instances
-
-### Tier 2: Spatial Questions (610 questions)
-- **proximity_percentage**: Percentage of one class near another
-- **proximity_area**: Area of one class near another
-- **binary_proximity**: Presence of one class near another
-- **building_proximity**: Number of buildings near other features
-- **building_flood_risk**: Buildings at flood risk (near water)
-- **building_fire_risk**: Buildings at fire risk (near forest)
-- **connectivity**: Counting isolated patches by size
-- **fragmentation**: Assessing if regions are connected or fragmented
-- **power_calculation**: Calculating solar panel power output
-
-### Tier 3: Complex Questions (677 questions)
-- **complex_multi_condition**: Areas meeting multiple spatial criteria
-- **complex_urban_flood_risk**: Urban areas at flood risk (near water)
-- **complex_urban_fire_risk**: Urban areas at fire risk (near forest)
-- **complex_agriculture_water_access**: Agricultural land with irrigation potential
-- **complex_size_filter**: Filtering by size thresholds
-- **complex_average**: Average sizes of regions
-
-## Loading the Dataset
+## Usage
 
 ```python
 from datasets import load_dataset
 
-# Load dataset
-dataset = load_dataset("PeterAM4/SQuID")
-
-# Access a sample
-sample = dataset['train'][0]
-image = sample['image']  # PIL Image
-question = sample['question']
-answer = sample['answer']  # String or numeric
-type = sample['type']
-
-# Convert answer based on type
-if type in ['percentage', 'count', 'proximity_percentage', 'proximity_area',
-            'building_proximity', 'building_flood_risk', 'building_fire_risk',
-            'connectivity', 'size', 'total_area', 'power_calculation'] or 'complex' in type:
-    answer_value = float(answer)
-elif 'binary' in type:
-    answer_value = int(answer)  # 0 or 1
-elif type == 'fragmentation':
-    answer_value = answer  # "connected" or "fragmented"
+bench = load_dataset("PeterAM4/SQuID", "benchmark", split="test")   # 2,000 questions, images + masks
+train = load_dataset("PeterAM4/SQuID", "train", split="train")      # 75,687 questions, file paths
 ```
 
-## Fields
-
-- **id**: Question identifier (e.g., "SQuID_0001")
-- **image**: Satellite image path
-- **question**: Question text with GSD notation
-- **answer**: Ground truth answer
-- **type**: One of 25 question types
-- **tier**: Difficulty level (1=Basic, 2=Spatial, 3=Complex)
-- **gsd**: Ground sampling distance in meters
-- **acceptable_range**: [lower, upper] bounds for numeric questions (when applicable)
+A numeric answer is correct if it falls in `[acceptable_range_lower, acceptable_range_upper]`. Binary (`0`/`1`) and `connected`/`fragmented` answers must match exactly.
 
-## Evaluation
-
-For numeric questions, check if predictions fall within the acceptable range:
+Training images and masks are stored as files. Download everything, including the raw JSON files used by the [code](https://github.com/PeterAMassih/qvlm-squid), with:
 
 ```python
-import math
-
-def evaluate(prediction, sample):
-    if 'acceptable_range' in sample:
-        # Numeric question - check if within human agreement range
-        lower, upper = sample['acceptable_range']
-        return lower <= float(prediction) <= upper
-    else:
-        # Non-numeric question - exact match required
-        return str(prediction).lower() == str(sample['answer']).lower()
+from huggingface_hub import snapshot_download
+snapshot_download("PeterAM4/SQuID", repo_type="dataset", local_dir="data",
+                  allow_patterns=["benchmark/*", "train/*"])
 ```
 
-The acceptable ranges represent the natural variation in human perception for spatial measurements.
-
-## Dataset Distribution
+## Data
 
-### By Tier
-- **Tier 1 (Basic)**: 713 questions (35.6%)
-- **Tier 2 (Spatial)**: 610 questions (30.5%)
-- **Tier 3 (Complex)**: 677 questions (33.9%)
+| Split | Questions | Images | Sources (GSD) |
+|---|---:|---:|---|
+| `benchmark` | 2,000 | 639 | DeepGlobe (0.5 m) 549 · EarthVQA (0.3 m) 1,304 · Solar (0.3 m) 97 · NAIP (1.0 m) 50 |
+| `train` | 75,687 | 1,443 | DeepGlobe 35,542 · EarthVQA 36,190 · Solar 3,955 |
 
-### Top Question Types
-- **complex_multi_condition**: 473 questions (23.6%)
-- **count**: 178 questions (8.9%)
-- **percentage**: 176 questions (8.8%)
-- **binary_comparison**: 175 questions (8.8%)
-- **binary_proximity**: 142 questions (7.1%)
-- **region_percentage**: 134 questions (6.7%)
-- **proximity_percentage**: 115 questions (5.8%)
-- **fragmentation**: 108 questions (5.4%)
-- **complex_agriculture_water_access**: 105 questions (5.2%)
-- **proximity_area**: 105 questions (5.2%)
+- **Tiers:** 1 basic (coverage, counts, sizes), 2 spatial (proximity, connectivity, power), 3 multi-step (filter by size, then clip by distance). The benchmark has 713 / 610 / 677 questions per tier and 25 question types.
+- **Splits** are image-disjoint and come from the official train / evaluation partitions of each source. The 50 NAIP questions are answered by humans (median of 10 annotators).
+- **Tolerances** come from the disagreement of 500 human annotations: ±1.735 points for percentages, ±2.25 for proximity percentages, ±max(1, 0.19·C) for counts, and ±1.735 % (±2.25 % near features) of the value for areas.
 
-### By Source
-- **DeepGlobe (0.5m GSD)**: 549 questions, 172 images - Land use classification masks
-- **EarthVQA (0.3m GSD)**: 1304 questions, 416 images - Building detection and land cover masks
-- **Solar Panels (0.3m GSD)**: 97 questions, 37 images - Solar panel segmentation masks
-- **NAIP (1.0m GSD)**: 50 questions, 14 images - Human-annotated diverse scenes
+**Fields.** `benchmark`: `id`, `image`, `mask` (none for NAIP), `question`, `answer` (string), `type`, `tier`, `gsd`, `acceptable_range_lower`, `acceptable_range_upper`, `source`. `train`: the same text fields, with `image` and `mask` as paths into `train/images_and_masks/`.
 
+**Masks.** DeepGlobe `*_label.png` and EarthVQA `*_label2.png` are RGB class maps, and Solar `*_label.bmp` is binary. EarthVQA's human labels leave about a quarter of each tile as background. `_label2` fills that background with a land-cover model's predictions (`_pred2`), and the original label is kept as `_label`.
 
-## Statistics Summary
+## Versions
 
-- **Zero-valued answers**: 82 (4.1%)
-- **Questions with ranges**: 1544 (77.2%)
-- **Average questions per image**: 3.1
+| Version | Content | |
+|---|---|---|
+| **v2.0** (current) | Benchmark + training split + ground-truth masks. Areas are exact pixel counts of connected components (holes excluded), vegetation is defined in the question text, ids are content hashes, and degenerate questions with a forced zero answer are removed. | `revision="v2.0"` |
+| v1.0 | Original benchmark (January 2026) | `revision="v1.0"`, [DOI 10.57967/hf/7565](https://doi.org/10.57967/hf/7565) |
 
-## Notes
+## License and citation
 
-- Questions explicitly state minimum area thresholds (e.g., "ignore patches smaller than 0.125 hectares")
-- Zero-valued answers indicate absence of features (intentionally included for robustness testing)
-- The benchmark tests both presence and absence of spatial features to avoid positive-only bias
-- Human agreement ranges allow for natural variation in spatial perception and counting
-- All measurements use metric units based on the specified GSD (Ground Sampling Distance)
-- Count ranges use proportional MADc (0.19) so larger counts have wider acceptable ranges
-
-## Citation
-
-If you use this dataset, please cite:
+Questions, answers and annotations are released under CC BY-NC 4.0. Images and masks are redistributed for non-commercial research from [DeepGlobe](https://arxiv.org/abs/1805.06561), [EarthVQA](https://github.com/Junjue-Wang/EarthVQA), the [PV segmentation dataset](https://doi.org/10.5281/zenodo.5171712) and [NAIP](https://naip-usdaonline.hub.arcgis.com/), and remain subject to their original terms.
 
 ```bibtex
-@article{massih2025squid,
-  title={Preserving Pixel-Level Precision: SQuID Dataset and QVLM Architecture for Quantitative Geospatial Reasoning},
-  author={Peter A. Massih, Eric Cosatto},
-  journal={arXiv preprint arXiv:XXXX.XXXXX},
-  year=2025
+@misc{massih2026reasoningpixellevelprecisionqvlm,
+  title         = {Reasoning with Pixel-level Precision: QVLM Architecture and SQuID Dataset for Quantitative Geospatial Analytics},
+  author        = {Peter A. Massih and Eric Cosatto},
+  year          = {2026},
+  eprint        = {2601.13401},
+  archivePrefix = {arXiv},
+  primaryClass  = {cs.CV},
+  url           = {https://arxiv.org/abs/2601.13401}
 }
 ```
-
----
-*Generated on 2026-07-10 14:39:57*