data: add released v23 holdout datasets to public registry (18 mirrors)

src/gasbench/dataset/config.py

@@ -25,7 +25,7 @@
 BENCHMARK_TOTAL_OVERRIDES = {
     "debug": {"image": 100, "video": 50, "audio": 50},
     "small": {"image": 5000, "video": 5000, "audio": 5000},
-    "full":  {"image": 69000, "video": 33000, "audio": 47000},
+    "full":  {"image": 72500, "video": 34000, "audio": 47500},
 }
 
 # Per-dataset download limits (only applied in debug/small modes for faster testing)

src/gasbench/dataset/configs/real_images.yaml

@@ -1151,3 +1151,75 @@ datasets:
     generator_family: real
     content_category: documents
     notes: Released from v22 bmcore holdouts
+
+  # ── Released v23 holdouts ─────────────────────────────────────────────────
+
+  - name: kidney-stone-ct
+    path: 34data/kidney-stone-ct
+    modality: image
+    media_type: real
+    source_format: parquet
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    generator_family: real
+    content_category: medical
+
+  - name: malaria-bounding-boxes
+    path: 34data/malaria-bounding-boxes
+    modality: image
+    media_type: real
+    source_format: parquet
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    generator_family: real
+    content_category: medical
+
+  - name: pulmonary-chest-xray-abnormalities
+    path: 34data/pulmonary-chest-xray-abnormalities
+    modality: image
+    media_type: real
+    source_format: parquet
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    generator_family: real
+    content_category: medical
+
+  - name: dermnet
+    path: 34data/dermnet
+    modality: image
+    media_type: real
+    source_format: parquet
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    generator_family: real
+    content_category: medical
+
+  - name: coronahack-chest-xray
+    path: 34data/coronahack-chest-xray
+    modality: image
+    media_type: real
+    source_format: parquet
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    generator_family: real
+    content_category: medical
+
+  - name: lits-256
+    path: 34data/lits-256
+    modality: image
+    media_type: real
+    source_format: parquet
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    generator_family: real
+    content_category: medical
+
+  - name: tcga-coad-msi-mss
+    path: 34data/tcga-coad-msi-mss
+    modality: image
+    media_type: real
+    source_format: parquet
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    generator_family: real
+    content_category: medical

src/gasbench/dataset/configs/synthetic_audio.yaml

@@ -796,3 +796,15 @@ datasets:
     generator_family: mixed-generation
     content_category: speech
     notes: Synthetic subset of the official RuASD anti-spoofing dataset
+
+  # ── Released v23 holdouts ─────────────────────────────────────────────────
+
+  - name: indic-tts-urdu
+    path: 34data/indic-tts-urdu
+    modality: audio
+    media_type: synthetic
+    source_format: zip
+    media_per_archive: 500
+    archives_per_dataset: 2
+    generator_family: mixed-generation
+    content_category: speech

src/gasbench/dataset/configs/synthetic_images.yaml

@@ -942,3 +942,35 @@ datasets:
     notes: PICA-100K is purely synthetic (PICABench; video-derived src and edited tgt). Not split — src is not camera-real.
     generator_family: mixed-generation
     content_category: diverse
+
+  # ── Released v23 holdouts ─────────────────────────────────────────────────
+
+  - name: synthetic-human
+    path: 34data/synthetic-human
+    modality: image
+    media_type: synthetic
+    source_format: parquet
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    generator_family: mixed-generation
+    content_category: faces
+
+  - name: ai-generated-ecommerce-damaged-product
+    path: 34data/ai-generated-ecommerce-damaged-product
+    modality: image
+    media_type: synthetic
+    source_format: parquet
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    generator_family: mixed-generation
+    content_category: diverse
+
+  - name: ai-generated-ecommerce-fake-logistics
+    path: 34data/ai-generated-ecommerce-fake-logistics
+    modality: image
+    media_type: synthetic
+    source_format: parquet
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    generator_family: mixed-generation
+    content_category: diverse

src/gasbench/dataset/configs/synthetic_videos.yaml

@@ -1266,3 +1266,75 @@ datasets:
     generator_family: mixed-generation
     content_category: diverse
     notes: Released from v22 bmcore holdouts
+
+  # ── Released v23 holdouts ─────────────────────────────────────────────────
+
+  - name: videogen-rewardbench-easyanimatev4
+    path: 34data/videogen-rewardbench-easyanimatev4
+    modality: video
+    media_type: synthetic
+    source_format: mp4
+    media_per_archive: 500
+    archives_per_dataset: 1
+    generator_family: easyanimate
+    content_category: diverse
+
+  - name: videogen-rewardbench-opensora1-2
+    path: 34data/videogen-rewardbench-opensora1-2
+    modality: video
+    media_type: synthetic
+    source_format: mp4
+    media_per_archive: 500
+    archives_per_dataset: 1
+    generator_family: opensora
+    content_category: diverse
+
+  - name: videogen-rewardbench-qingying
+    path: 34data/videogen-rewardbench-qingying
+    modality: video
+    media_type: synthetic
+    source_format: mp4
+    media_per_archive: 500
+    archives_per_dataset: 1
+    generator_family: qingying
+    content_category: diverse
+
+  - name: videogen-rewardbench-tongyi
+    path: 34data/videogen-rewardbench-tongyi
+    modality: video
+    media_type: synthetic
+    source_format: mp4
+    media_per_archive: 500
+    archives_per_dataset: 1
+    generator_family: tongyi
+    content_category: diverse
+
+  - name: videogen-rewardbench-vidu
+    path: 34data/videogen-rewardbench-vidu
+    modality: video
+    media_type: synthetic
+    source_format: mp4
+    media_per_archive: 500
+    archives_per_dataset: 1
+    generator_family: vidu
+    content_category: diverse
+
+  - name: edtalk
+    path: 34data/edtalk
+    modality: video
+    media_type: semisynthetic
+    source_format: mp4
+    media_per_archive: 500
+    archives_per_dataset: 1
+    generator_family: edtalk
+    content_category: faces
+
+  - name: float-talking-head
+    path: 34data/float-talking-head
+    modality: video
+    media_type: semisynthetic
+    source_format: mp4
+    media_per_archive: 500
+    archives_per_dataset: 1
+    generator_family: float
+    content_category: faces