Release v23 holdout datasets to the public registry

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "gasbench"
-version = "0.9.3"
+version = "0.9.4"
 description = "GASBench - ML model benchmark evaluation package"
 readme = "README.md"
 requires-python = ">=3.10"

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_audio.yaml

@@ -755,3 +755,57 @@ datasets:
     generator_family: real
     content_category: speech
     notes: CML-TTS Spanish train split, released from v22 bmcore holdouts
+
+  # ── Released v23 holdouts — deferred to upstream repos (no 34data mirror) ──
+
+  - name: cml-tts-dutch
+    path: ylacombe/cml-tts
+    hf_revision: 48aedd7a8ea08efc75cae3e81fbefa9720df0812
+    modality: audio
+    media_type: real
+    source_format: parquet
+    media_per_archive: 500
+    archives_per_dataset: 2
+    hf_subfolders: ['dutch']
+    include_paths: ['train-']
+    generator_family: real
+    content_category: speech
+
+  - name: cml-tts-italian
+    path: ylacombe/cml-tts
+    hf_revision: 48aedd7a8ea08efc75cae3e81fbefa9720df0812
+    modality: audio
+    media_type: real
+    source_format: parquet
+    media_per_archive: 500
+    archives_per_dataset: 2
+    hf_subfolders: ['italian']
+    include_paths: ['train-']
+    generator_family: real
+    content_category: speech
+
+  - name: cml-tts-polish
+    path: ylacombe/cml-tts
+    hf_revision: 48aedd7a8ea08efc75cae3e81fbefa9720df0812
+    modality: audio
+    media_type: real
+    source_format: parquet
+    media_per_archive: 500
+    archives_per_dataset: 2
+    hf_subfolders: ['polish']
+    include_paths: ['train-']
+    generator_family: real
+    content_category: speech
+
+  - name: cml-tts-portuguese
+    path: ylacombe/cml-tts
+    hf_revision: 48aedd7a8ea08efc75cae3e81fbefa9720df0812
+    modality: audio
+    media_type: real
+    source_format: parquet
+    media_per_archive: 500
+    archives_per_dataset: 2
+    hf_subfolders: ['portuguese']
+    include_paths: ['train-']
+    generator_family: real
+    content_category: speech

src/gasbench/dataset/configs/real_images.yaml

@@ -1151,3 +1151,102 @@ 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
+
+  # ── Released v23 holdouts — deferred to upstream repos (no 34data mirror) ──
+
+  - name: conceptual-captions-0-1m
+    path: obscure-entropy/conceptual_captions_hu_filtered
+    hf_revision: e70871e77287ca4aec4a12ff2f9cec32cad5694a
+    modality: image
+    media_type: real
+    source_format: parquet
+    data_columns: [img]
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    include_paths: ['data/conceptual_captions_0_1m.parquet']
+    generator_family: real
+    content_category: diverse
+
+  - name: rocov2-radiology
+    path: eltorio/ROCOv2-radiology
+    hf_revision: 80ffeef4eb8d34d27cb5c2815305f1d8aee8a83c
+    modality: image
+    media_type: real
+    source_format: parquet
+    media_per_archive: 1000
+    archives_per_dataset: 1
+    include_paths: ['data/train-']
+    generator_family: real
+    content_category: medical

src/gasbench/dataset/configs/real_videos.yaml

@@ -1093,3 +1093,65 @@ datasets:
     generator_family: real
     content_category: action
     notes: Released from v22 bmcore holdouts
+
+  # ── Released v23 holdouts — deferred to upstream repos (no 34data mirror) ──
+
+  - name: timelens2-93k-00003
+    path: MCG-NJU/TimeLens2-93K
+    hf_revision: fce4049cdf5d720e2651b956c58e664c3386c021
+    modality: video
+    media_type: real
+    source_format: tar
+    media_per_archive: 500
+    archives_per_dataset: 1
+    include_paths: ['videos/videos-00003-of-00018.tar']
+    generator_family: real
+    content_category: diverse
+
+  - name: llava-onevision-2-mid-training-00000
+    path: mvp-lab/LLaVA-OneVision-2-Data
+    hf_revision: bf7038ba5d17771dc113a3072f98591ad2e29541
+    modality: video
+    media_type: real
+    source_format: tar
+    media_per_archive: 500
+    archives_per_dataset: 1
+    include_paths: ['mid_training_video/30s/train_00000_of_09707.tar']
+    generator_family: real
+    content_category: diverse
+
+  - name: rekadaily-phone-shard-00000
+    path: RekaAI/RekaDaily-10k-raw
+    hf_revision: a42e2da9aaeef7c8653d9de50772c97ceb954251
+    modality: video
+    media_type: real
+    source_format: tar
+    media_per_archive: 500
+    archives_per_dataset: 1
+    include_paths: ['data/video_capture_first_person_videos_phone/shard-00000.tar']
+    generator_family: real
+    content_category: action
+
+  - name: rekadaily-phone-shard-00001
+    path: RekaAI/RekaDaily-10k-raw
+    hf_revision: a42e2da9aaeef7c8653d9de50772c97ceb954251
+    modality: video
+    media_type: real
+    source_format: tar
+    media_per_archive: 500
+    archives_per_dataset: 1
+    include_paths: ['data/video_capture_first_person_videos_phone/shard-00001.tar']
+    generator_family: real
+    content_category: action
+
+  - name: rekadaily-phone-shard-00002
+    path: RekaAI/RekaDaily-10k-raw
+    hf_revision: a42e2da9aaeef7c8653d9de50772c97ceb954251
+    modality: video
+    media_type: real
+    source_format: tar
+    media_per_archive: 500
+    archives_per_dataset: 1
+    include_paths: ['data/video_capture_first_person_videos_phone/shard-00002.tar']
+    generator_family: real
+    content_category: action

src/gasbench/dataset/configs/synthetic_audio.yaml

@@ -796,3 +796,67 @@ 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
+
+  - name: ai4bharat-speecharenabench-gu
+    path: ai4bharat/SpeechArenaBench
+    modality: audio
+    media_type: synthetic
+    source_format: parquet
+    media_per_archive: 500
+    archives_per_dataset: 2
+    hf_revision: fdd26e85ff051036e3eb49d7053d2b8f15c14ec0
+    hf_subfolders: [gu]
+    data_columns: [audio_a]
+    generator_family: mixed-generation
+    content_category: speech
+
+  - name: ai4bharat-speecharenabench-ml
+    path: ai4bharat/SpeechArenaBench
+    modality: audio
+    media_type: synthetic
+    source_format: parquet
+    media_per_archive: 500
+    archives_per_dataset: 2
+    hf_revision: fdd26e85ff051036e3eb49d7053d2b8f15c14ec0
+    hf_subfolders: [ml]
+    data_columns: [audio_a]
+    generator_family: mixed-generation
+    content_category: speech
+
+  - name: ai4bharat-speecharenabench-or
+    path: ai4bharat/SpeechArenaBench
+    modality: audio
+    media_type: synthetic
+    source_format: parquet
+    media_per_archive: 500
+    archives_per_dataset: 2
+    hf_revision: fdd26e85ff051036e3eb49d7053d2b8f15c14ec0
+    hf_subfolders: [or]
+    data_columns: [audio_a]
+    generator_family: mixed-generation
+    content_category: speech
+
+  - name: ai4bharat-speecharenabench-ur
+    path: ai4bharat/SpeechArenaBench
+    modality: audio
+    media_type: synthetic
+    source_format: parquet
+    media_per_archive: 500
+    archives_per_dataset: 2
+    hf_revision: fdd26e85ff051036e3eb49d7053d2b8f15c14ec0
+    hf_subfolders: [ur]
+    data_columns: [audio_a]
+    generator_family: mixed-generation
+    content_category: speech