Upload folder using huggingface_hub

README.md

@@ -0,0 +1,153 @@
+---
+configs:
+  - config_name: all
+    default: true
+    data_files:
+      - split: train
+        path: data/all/train.parquet
+      - split: test
+        path: data/all/test.parquet
+  - config_name: qrr
+    data_files:
+      - split: train
+        path: data/qrr/train.parquet
+      - split: test
+        path: data/qrr/test.parquet
+  - config_name: trr
+    data_files:
+      - split: train
+        path: data/trr/train.parquet
+      - split: test
+        path: data/trr/test.parquet
+  - config_name: fdr
+    data_files:
+      - split: train
+        path: data/fdr/train.parquet
+      - split: test
+        path: data/fdr/test.parquet
+  - config_name: multiview
+    data_files:
+      - split: train
+        path: data/multiview/train.parquet
+      - split: test
+        path: data/multiview/test.parquet
+task_categories:
+  - visual-question-answering
+language:
+  - en
+license: mit
+tags:
+  - spatial-reasoning
+  - vlm-benchmark
+  - ordinal-relations
+  - 3d-scenes
+  - multi-view
+size_categories:
+  - 100K<n<1M
+---
+
+# ORDINARY-BENCH
+
+A benchmark for evaluating Vision-Language Models (VLMs) on **ordinal spatial reasoning** in 3D scenes.
+
+## Overview
+
+- **700 synthetic 3D scenes** rendered with Blender (CLEVR-style)
+- **7 complexity levels**: 4 to 10 objects per scene (100 scenes each)
+- **332,857 questions** across 3 reasoning types
+- **Single-view and multi-view** (4 camera angles) images
+
+## Question Types
+
+### QRR (Quantitative Relation Reasoning) — 130,557 questions
+Compare 3D distances between object pairs.
+- **Disjoint**: `dist(A,B)` vs `dist(C,D)` — answer `<`, `~=`, or `>`
+- **Shared anchor**: `dist(A,B)` vs `dist(A,C)` — common anchor object
+
+### TRR (Ternary Relation Reasoning) — 197,400 questions
+Determine clock-face direction: standing at `ref1`, facing `ref2` (12 o'clock), what hour (1-12) is `target` at?
+
+### FDR (Full Distance Ranking) — 4,900 questions
+Rank all other objects by 3D distance from an anchor, nearest to farthest.
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+# Load all questions with single-view images
+ds = load_dataset("./datasets", "all", split="test")
+
+# Load specific question type
+ds_qrr = load_dataset("./datasets", "qrr", split="test")
+ds_trr = load_dataset("./datasets", "trr", split="test")
+ds_fdr = load_dataset("./datasets", "fdr", split="test")
+
+# Load with multi-view images (4 views per scene)
+ds_mv = load_dataset("./datasets", "multiview", split="test")
+
+# Access a sample
+sample = ds_qrr[0]
+print(sample["question_text"])    # Natural language question
+print(sample["qrr_gt_comparator"])  # Ground truth: "<", "~=", or ">"
+sample["image"].show()            # Display the scene image
+```
+
+## Configs
+
+| Config | Description | Image columns |
+|--------|-------------|---------------|
+| `all` (default) | All 3 question types | `image` (single-view) |
+| `qrr` | QRR questions only | `image` |
+| `trr` | TRR questions only | `image` |
+| `fdr` | FDR questions only | `image` |
+| `multiview` | All types, 4 views | `image`, `image_view0`..`image_view3` |
+
+## Data Splits
+
+| Split | Scenes per complexity | Total scenes | Description |
+|-------|----------------------|--------------|-------------|
+| train | 80 | 560 | Scene indices 000-079 |
+| test | 20 | 140 | Scene indices 080-099 |
+
+## Column Schema
+
+**Common columns**: `scene_id`, `n_objects`, `split`, `image`, `objects` (JSON), `question_type`, `qid`, `question_text`, `scene_metadata` (JSON)
+
+**QRR-specific**: `qrr_variant`, `qrr_pair1`, `qrr_pair2`, `qrr_metric`, `qrr_gt_comparator`
+
+**TRR-specific**: `trr_target`, `trr_ref1`, `trr_ref2`, `trr_gt_hour`, `trr_gt_quadrant`, `trr_gt_angle_deg`
+
+**FDR-specific**: `fdr_anchor`, `fdr_n_ranked`, `fdr_gt_ranking`, `fdr_gt_distances`, `fdr_gt_tie_groups`
+
+## Scoring
+
+- **QRR**: Exact comparator match (`<`, `~=`, `>`)
+- **TRR**: Three granularities — hour (exact), quadrant (same 1/4), adjacent (within ±1 hour)
+- **FDR**: Four metrics — exact match, Kendall τ, pairwise accuracy, top-1 accuracy
+
+## Prompt Templates
+
+System prompts for VLM evaluation are included in `prompts/system_prompts.json`. These define the expected response format (JSON array of `{qid, answer}` objects).
+
+## Building the Dataset
+
+The dataset is built from raw generation outputs using:
+
+```bash
+cd datasets
+pip install pandas pyarrow datasets pillow
+python build_dataset.py
+```
+
+This copies images and generates Parquet files from scene JSONs and question files.
+
+Options:
+- `--dry-run` — print statistics without writing files
+- `--skip-images` — skip image copying (faster for development)
+- `--data-gen-dir PATH` — custom data-gen output directory
+- `--questions-dir PATH` — custom questions directory
+
+## License
+
+MIT

data/all/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba6356110d041e2a44efb7f926cb72eea8e1f6fde08a7ecf462b958799b5083c
+size 2615914

data/all/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b5210eb457bb5a274c72f7cbc3bd9fd727f3973e18fefcf437aca87971177e5
+size 10374044

data/fdr/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ca86261b099e1d46fcb6a7bd6ac6c0b73706c7db118027c407983778090c5efc
+size 340270

data/fdr/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f797b70cd6e9fb780c098db593e7d357425df1521b74c70299bab2545abe8e90
+size 1347500

data/multiview/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aee02741511a1028c8f8053f0bfe3d1b83be16a2b0a6451814200df8c4ffa41f
+size 7349457

data/multiview/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6fe3e157d925d72ec7b35bf52d9dddb8ab7293ddbc56e9671034b4fb2ac568c5
+size 29051628

data/qrr/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7f78f23fafec239d1a48f0c31f1e867287d516b609b4a07e432f6102f2a39e69
+size 869034

data/qrr/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9a0284eeb7d2b0483571a370f13e9259ad63af4ac0429e7ebb462f1e785c661
+size 3424893

data/trr/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43a62da2c0da1002386c62886ec319044e629e2e680c2a05177647c698531a11
+size 1373168

data/trr/train.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b46f6bf3308c33c665ad21dcf359e8506f1dd94596e3ef2126b75a337bf0ed4b
+size 5455011