Fix Dataset Viewer train split configuration

Explicitly map the default train split to data/train/**, use metadata.jsonl plus images/, prevent scientific validation JSON from being inferred as a split, and remove redundant embedded-image Parquet Viewer shards.

.gitignore

@@ -20,3 +20,7 @@ viewer_data/
 Krea-2-Turbo-Viewer-Fix-UPLOAD.zip
 viewer_fix_*.log
 viewer_server_diagnostics.json
+
+# Local Dataset Viewer repair artifacts
+imagefolder_viewer_*.log
+imagefolder_viewer_diagnostics.json

README.md

@@ -8,7 +8,9 @@ size_categories:
 configs:
 - config_name: default
   default: true
-  data_dir: data/train
+  data_files:
+  - split: train
+    path: "data/train/**"
   drop_labels: true
 task_categories:
 - text-to-image
@@ -29,7 +31,7 @@ tags:
 - q8-0
 - q4-k-m
 ---
-<!-- Dataset Viewer repair: the Viewer reads the small Hugging Face-native Parquet shards in `viewer_data/train-*.parquet` (240 rows, 8 shards). The original ImageFolder and scientific artifacts remain available separately. -->
+<!-- Dataset Viewer source: the explicit train split reads data/train/metadata.jsonl and data/train/images/ with Hugging Face ImageFolder. -->
 # Krea 2 Turbo ComfyUI Format Fidelity Benchmark
 
 This release is a paired, deterministic comparison of eight Krea 2 Turbo checkpoint formats in ComfyUI: BF16, FP8 Scaled, INT8 ConvRot, MXFP8, NVFP4, INT4 ConvRot W4A4, GGUF Q8_0, and GGUF Q4_K_M. It contains 240 scored 1024×1024 images, saved float32 decoded tensors and final latents, every denoising trajectory, raw metric tables, telemetry, statistical comparisons, and reproduction code.
@@ -56,10 +58,8 @@ The test GPU was an NVIDIA GeForce RTX 4060 Ti 16 GB (SM 8.9). MXFP8 and NVFP4 n
 The original five formats and the three additions were measured in separate sessions. Five unscored BF16 and INT8 ConvRot bridge repeats quantify drift. INT8 crossed the preregistered 5% drift threshold, so exact old-versus-new timing comparisons remain caveated; fidelity comparisons remain paired to the same saved BF16 prompt/seed outputs.
 
 ## Dataset organization
-  * `data/train/metadata.jsonl` and `data/train/images/` are the Dataset Viewer and loading source, using Hugging Face ImageFolder. Each metadata row links an image to its prompt, seed, format, checkpoint provenance, raw scientific artifacts, and flattened metric values.
 
-- `viewer_data/train-*.parquet` are the Dataset Viewer source: small, self-contained Parquet shards with the embedded image plus every prompt, seed, format, checkpoint provenance, raw-artifact path, and flattened metric column.
-- `data/train/metadata.jsonl` and `data/train/images/` are the equivalent ImageFolder tree kept for direct file access. Each metadata row links an image to its prompt, seed, format, checkpoint provenance, raw scientific artifacts, and flattened metric values.
+- `data/train/metadata.jsonl` and `data/train/images/` are the Dataset Viewer and loading source, using Hugging Face ImageFolder. The dataset card explicitly limits the `train` split to this tree so scientific JSON files elsewhere in the repository are not inferred as dataset splits. Each metadata row links an image to its prompt, seed, format, checkpoint provenance, raw scientific artifacts, and flattened metric values.
 - `raw/` contains `decoded_float32.npy`, `final_latent_float32.npy`, `trajectory.npz`, and capture `metadata.json` for every scored run.
 - `metrics/` contains raw per-image, per-parameter, paired-statistics, summary, trajectory, latency, and performance tables.
 - `comparison_sheets/` contains eight-format sheets, BF16-relative difference maps, and automatically selected detail crops.

checksums/SHA256SUMS

@@ -1,9 +1,9 @@
 26dd9f691c984a8676e4ff0dccf50239d35c174e863d6884c3d43818b0ab577c  .gitattributes
-38d9694f003d651d39799317c174d65cb27784ac8042f60ad9189c138172353a  .gitignore
+52d93f44a3403e5e85615ce04b8371d60ff079867ac700f3f134107acc1244e1  .gitignore
 afbd7232877f7fb929d8e7a2b3b396a41179a3106ee6c1c99f90aa470062f969  CITATION.cff
 fe888a3acbf08a05c3e976f76aa635f2a7034d09589955117c890f3f3a05c0a6  LICENSE-CODE
 10232ffb753de2527c89b1f4a69bb1479e846892c3b3e26a8550866d0f0fe440  LICENSE-DATA
-59717bd9aeefec2b035370c1bde0c0dbb4dd5e5a5bdb1465bfac523b2259883b  README.md
+6497d1fda5a3d0b4dc0be66fe50b3253c2c35f0977f3eac87ab42301651688e9  README.md
 7860f02cff561fc62a93cabc452ea76a20452063e8207f97fecedb1a74fbadb2  TECHNICAL_REPORT.md
 f020293711e8ad7eab2758442e5479b39648cd6c725fd076403f66a029c60b5c  comparison_sheets/complete.json
 d71bfb4a773b3998636d68998ad885b52907b285a703cb2348976f4b613ba4de  comparison_sheets/contact_sheet_replicate0.png
@@ -98,7 +98,6 @@ a1495f17df7382e30c740db4d07574f34e5e31cb20020ac7fc9f71a4fea002bb  comparison_she
 b2705f64ed3fca264b0e7d6c6d3a01fa2f977477fd0a19ca94537552b5bba036  comparison_sheets/full/p15_scientific_poster__r0.png
 d3d58661e35a3b4b6b0c4fe11f6f09f2c66dab482c89f4264a872806e3ead10a  comparison_sheets/full/p15_scientific_poster__r1.png
 a4335e2c6da478e4eab3e2ccaa933553d0b4810179064893b57ed52001a1677a  comparison_sheets/index.json
-08b9c4cb964aed1251a1835eda6943397cdf94dac89e14161086761a743b94d6  data/train-00000-of-00001.parquet
 fd760338f2c01688676cd31a47421de131674706ae9d18b364bc5d5346ce6de2  data/train/images/bf16/p01_portrait__r0__bf16.png
 5f332fe49f9202be8ebc0d43f122fcf1725a384fcfa9d14a5caee0471f005e08  data/train/images/bf16/p01_portrait__r1__bf16.png
 55fb84113e54887184c8f216ac376cb6af3f311f2a44b4dbb8a24861576d5014  data/train/images/bf16/p02_hands_group__r0__bf16.png
@@ -1360,9 +1359,9 @@ e77eff0f644162b32fdbb19e0787b1381a007b41897a4bacb9b47c18791b6a4d  reproduction/b
 150a133f4ff5508ca9828fd436fd3db854fbc7218c8fcd9217395eabdb51e3b3  reproduction/download_models.py
 ac43fe8c1625b5754ab3d829591758a50607320f4f694c32784e5ce026c2bc59  reproduction/requirements-release.txt
 56ecc0611beb0a2ed147b3eed17da521b2c21c0394776082ced988ec8ae63dc7  scripts/build_parquet.py
-b393da3601e6cc985e4db4f392260ff26ea760e63a0d818bedf4ae7c39d82acf  scripts/prepare_release.py
-15a332d8bd9031c64539d7bd64df1f55619987ae9e07d06540585fc302d8ac82  scripts/upload_to_huggingface.py
-003d039c79e007f6d36cb46279ae2cd1d5509200b4c3ad13d2f3faa28651c878  scripts/validate_release.py
+d2e33bcf2ff59345414a00be97737776f0d46058abefb6a72e242ed2f89c1d94  scripts/prepare_release.py
+4e6a346c63f2d8ad0d00a7589177fbcb4b895c9c9df8cdef10d3dc26cd97a270  scripts/upload_to_huggingface.py
+0e75a48c4bbea75d36ce0ecce98b0db7a6552a9f77be14810c385745faca4792  scripts/validate_release.py
 99eb77d4339d532e6fc11cfd50cd55b4f8db54f58f1ddcd93cf74361a9252012  tables/decision_table.csv
 74f759089450fe7f074d67a389988251422d78260bdd728d040cf7fde7a365fb  tables/format_ranking.csv
 1434a8b74c4e55dc0dcbe7b0aac9fae34c4fc8fc088051724aaf3b139fc7392c  tables/metadata_schema.json

scripts/prepare_release.py

@@ -364,12 +364,13 @@ license: cc-by-4.0
 pretty_name: Krea 2 Turbo ComfyUI Format Fidelity Benchmark
 size_categories:
 - n<1K
-configs:
-- config_name: default
-  default: true
-  data_files:
-  - split: train
-    path: data/train-*.parquet
+configs:
+- config_name: default
+  default: true
+  data_files:
+  - split: train
+    path: "data/train/**"
+  drop_labels: true
 task_categories:
 - text-to-image
 tags:
@@ -388,7 +389,8 @@ tags:
 - gguf
 - q8-0
 - q4-k-m
----
+---
+<!-- Dataset Viewer source: the explicit train split reads data/train/metadata.jsonl and data/train/images/ with Hugging Face ImageFolder. -->
 
 # Krea 2 Turbo ComfyUI Format Fidelity Benchmark
 
@@ -415,10 +417,9 @@ The test GPU was an NVIDIA GeForce RTX 4060 Ti 16 GB (SM 8.9). MXFP8 and NVFP4 n
 
 The original five formats and the three additions were measured in separate sessions. Five unscored BF16 and INT8 ConvRot bridge repeats quantify drift. INT8 crossed the preregistered 5% drift threshold, so exact old-versus-new timing comparisons remain caveated; fidelity comparisons remain paired to the same saved BF16 prompt/seed outputs.
 
-## Dataset organization
-
-- `data/train-00000-of-00001.parquet` is the Dataset Viewer source: one self-contained table with the embedded image plus every prompt, seed, format, checkpoint provenance, raw-artifact path, and flattened metric column.
-- `data/train/metadata.jsonl` and `data/train/images/` are the equivalent ImageFolder tree kept for direct file access. Each metadata row links an image to its prompt, seed, format, checkpoint provenance, raw scientific artifacts, and flattened metric values.
+## Dataset organization
+
+- `data/train/metadata.jsonl` and `data/train/images/` are the Dataset Viewer and loading source, using Hugging Face ImageFolder. The dataset card explicitly limits the `train` split to this tree so scientific JSON files elsewhere in the repository are not inferred as dataset splits. Each metadata row links an image to its prompt, seed, format, checkpoint provenance, raw scientific artifacts, and flattened metric values.
 - `raw/` contains `decoded_float32.npy`, `final_latent_float32.npy`, `trajectory.npz`, and capture `metadata.json` for every scored run.
 - `metrics/` contains raw per-image, per-parameter, paired-statistics, summary, trajectory, latency, and performance tables.
 - `comparison_sheets/` contains eight-format sheets, BF16-relative difference maps, and automatically selected detail crops.
@@ -649,8 +650,10 @@ if __name__ == "__main__":
 def build_upload_script() -> str:
     return '''from __future__ import annotations
 
-import argparse
-from pathlib import Path
+import argparse
+import subprocess
+import sys
+from pathlib import Path
 
 from huggingface_hub import HfApi
 
@@ -665,8 +668,14 @@ def main() -> int:
     root = args.root.resolve()
     required = [root / "README.md", root / "data" / "train" / "metadata.jsonl", root / "checksums" / "SHA256SUMS"]
     missing = [str(path) for path in required if not path.is_file()]
-    if missing:
-        raise RuntimeError("Release is incomplete: " + ", ".join(missing))
+    if missing:
+        raise RuntimeError("Release is incomplete: " + ", ".join(missing))
+    if (root / "viewer_data").exists() or (root / "data" / "train-00000-of-00001.parquet").exists():
+        raise RuntimeError("Obsolete Dataset Viewer Parquet artifacts are present")
+    subprocess.run(
+        [sys.executable, str(root / "scripts" / "validate_release.py"), "--root", str(root)],
+        check=True,
+    )
     files = [path for path in root.rglob("*") if path.is_file()]
     total = sum(path.stat().st_size for path in files)
     print(f"ready: {len(files)} files, {total / 1024**3:.3f} GiB -> datasets/{args.repo_id}")
@@ -679,7 +688,15 @@ def main() -> int:
         repo_type="dataset",
         folder_path=str(root),
         private=args.private,
-        ignore_patterns=[".cache/**", "**/__pycache__/**", "*.pyc"],
+        ignore_patterns=[
+            ".cache/**",
+            ".viewer_fix_venv/**",
+            "viewer_data/**",
+            "viewer_fix_output/**",
+            "**/__pycache__/**",
+            "*.pyc",
+            "viewer_fix_*.log",
+        ],
         num_workers=4,
         print_report_every=30,
     )
@@ -692,29 +709,7 @@ if __name__ == "__main__":
 '''
 
 
-def build_image_parquet(destination: Path, row_group_size: int = 100) -> Path:
-    import pyarrow.parquet as pq
-    from datasets import load_dataset
-    from datasets.table import embed_table_storage
-
-    out = destination / "data" / "train-00000-of-00001.parquet"
-    dataset = load_dataset(str(destination), split="train")
-    ordered = ["image"] + [name for name in dataset.column_names if name != "image"]
-    dataset = dataset.select_columns(ordered)
-    table = embed_table_storage(dataset.with_format("arrow")[:])
-    writer = pq.ParquetWriter(str(out), table.schema)
-    try:
-        for batch in table.to_batches(max_chunksize=row_group_size):
-            writer.write_batch(batch, row_group_size=row_group_size)
-    finally:
-        writer.close()
-    first = pq.read_table(str(out)).column("image")[0].as_py()
-    if not (isinstance(first, dict) and first.get("bytes")):
-        raise RuntimeError("Parquet build did not embed image bytes")
-    return out
-
-
-def build_release(benchmark_root: Path, destination: Path, config_path: Path) -> None:
+def build_release(benchmark_root: Path, destination: Path, config_path: Path) -> None:
     benchmark_root = benchmark_root.resolve()
     config_path = config_path if config_path.is_absolute() else benchmark_root / config_path
     config = load_configuration(config_path)
@@ -797,8 +792,7 @@ def build_release(benchmark_root: Path, destination: Path, config_path: Path) ->
                 if key not in IDENTIFIER_COLUMNS:
                     row[f"{group}_{key}"] = coerce(value)
         metadata_rows.append(row)
-    jsonl_dump(destination / "data" / "train" / "metadata.jsonl", metadata_rows)
-    build_image_parquet(destination)
+    jsonl_dump(destination / "data" / "train" / "metadata.jsonl", metadata_rows)
 
     copy_tree(results / "metrics", destination / "metrics")
     copy_tree(results / "comparison_sheets", destination / "comparison_sheets")

scripts/upload_to_huggingface.py

@@ -1,7 +1,9 @@
 from __future__ import annotations
 
-import argparse
-from pathlib import Path
+import argparse
+import subprocess
+import sys
+from pathlib import Path
 
 from huggingface_hub import HfApi
 
@@ -16,8 +18,14 @@ def main() -> int:
     root = args.root.resolve()
     required = [root / "README.md", root / "data" / "train" / "metadata.jsonl", root / "checksums" / "SHA256SUMS"]
     missing = [str(path) for path in required if not path.is_file()]
-    if missing:
-        raise RuntimeError("Release is incomplete: " + ", ".join(missing))
+    if missing:
+        raise RuntimeError("Release is incomplete: " + ", ".join(missing))
+    if (root / "viewer_data").exists() or (root / "data" / "train-00000-of-00001.parquet").exists():
+        raise RuntimeError("Obsolete Dataset Viewer Parquet artifacts are present")
+    subprocess.run(
+        [sys.executable, str(root / "scripts" / "validate_release.py"), "--root", str(root)],
+        check=True,
+    )
     files = [path for path in root.rglob("*") if path.is_file()]
     total = sum(path.stat().st_size for path in files)
     print(f"ready: {len(files)} files, {total / 1024**3:.3f} GiB -> datasets/{args.repo_id}")
@@ -30,7 +38,15 @@ def main() -> int:
         repo_type="dataset",
         folder_path=str(root),
         private=args.private,
-        ignore_patterns=[".cache/**", "**/__pycache__/**", "*.pyc"],
+        ignore_patterns=[
+            ".cache/**",
+            ".viewer_fix_venv/**",
+            "viewer_data/**",
+            "viewer_fix_output/**",
+            "**/__pycache__/**",
+            "*.pyc",
+            "viewer_fix_*.log",
+        ],
         num_workers=4,
         print_report_every=30,
     )

scripts/validate_release.py

@@ -8,10 +8,19 @@ import math
 import re
 from collections import Counter
 from pathlib import Path
-from urllib.parse import unquote
+from urllib.parse import unquote
+
+import yaml
 
 
-FORMATS = ("bf16", "fp8_scaled", "int8_convrot", "mxfp8", "nvfp4", "int4_convrot", "gguf_q8_0", "gguf_q4_k_m")
+FORMATS = ("bf16", "fp8_scaled", "int8_convrot", "mxfp8", "nvfp4", "int4_convrot", "gguf_q8_0", "gguf_q4_k_m")
+IMAGEFOLDER_GLOB = "data/train/**"
+LOCAL_REPAIR_DIRS = {".viewer_fix_venv", "viewer_fix_output", "__pycache__"}
+LOCAL_REPAIR_FILES = {
+    "imagefolder_viewer_diagnostics.json",
+    "imagefolder_viewer_fix.log",
+    "viewer_server_diagnostics.json",
+}
 
 
 def strict_json_loads(payload: str) -> object:
@@ -36,9 +45,59 @@ def checksum(path: Path) -> str:
     return value.hexdigest()
 
 
-def fail(condition: bool, message: str, errors: list[str]) -> None:
-    if condition:
-        errors.append(message)
+def fail(condition: bool, message: str, errors: list[str]) -> None:
+    if condition:
+        errors.append(message)
+
+
+def validate_dataset_card(root: Path, errors: list[str]) -> None:
+    readme = (root / "README.md").read_text(encoding="utf-8")
+    match = re.match(r"\A---\s*\n(.*?)\n---\s*\n", readme, re.DOTALL)
+    if not match:
+        errors.append("README.md has no valid YAML front matter")
+        return
+
+    try:
+        metadata = yaml.safe_load(match.group(1)) or {}
+    except yaml.YAMLError as exc:
+        errors.append(f"README.md YAML is invalid: {exc}")
+        return
+
+    configs = metadata.get("configs")
+    if not isinstance(configs, list):
+        errors.append("README.md must define a configs list")
+        return
+
+    defaults = [
+        config
+        for config in configs
+        if isinstance(config, dict) and config.get("config_name") == "default"
+    ]
+    if len(defaults) != 1:
+        errors.append("README.md must define exactly one default config")
+        return
+
+    config = defaults[0]
+    fail(config.get("default") is not True, "default config is not marked default", errors)
+    fail(config.get("drop_labels") is not True, "default config must set drop_labels: true", errors)
+    fail("data_dir" in config, "default config must not use data_dir", errors)
+
+    data_files = config.get("data_files")
+    expected = [{"split": "train", "path": IMAGEFOLDER_GLOB}]
+    fail(
+        data_files != expected,
+        f"default config data_files must equal {expected!r}, found {data_files!r}",
+        errors,
+    )
+
+
+def is_local_repair_artifact(root: Path, path: Path) -> bool:
+    relative = path.relative_to(root)
+    return (
+        any(part in LOCAL_REPAIR_DIRS for part in relative.parts)
+        or path.name in LOCAL_REPAIR_FILES
+        or (path.name.startswith("viewer_fix_") and path.suffix == ".log")
+    )
 
 
 def main() -> int:
@@ -46,8 +105,9 @@ def main() -> int:
     parser.add_argument("--root", type=Path, default=Path(__file__).resolve().parents[1])
     parser.add_argument("--full", action="store_true", help="Decode all images, inspect all arrays, verify all SHA-256 values, and load ImageFolder")
     args = parser.parse_args()
-    root = args.root.resolve()
-    errors: list[str] = []
+    root = args.root.resolve()
+    errors: list[str] = []
+    validate_dataset_card(root, errors)
     metadata_path = root / "data" / "train" / "metadata.jsonl"
     rows = []
     for line_number, line in enumerate(metadata_path.read_text(encoding="utf-8").splitlines(), start=1):
@@ -77,8 +137,9 @@ def main() -> int:
             required_paths.append(path)
             fail(not path.is_file(), f"missing {key}: {path}", errors)
     fail(len(list((root / "data" / "train" / "images").rglob("*.png"))) != expected_rows, f"expected {expected_rows} PNG images", errors)
-    parquet_path = root / "data" / "train-00000-of-00001.parquet"
-    fail(not parquet_path.is_file(), f"missing viewer parquet: {parquet_path}", errors)
+    parquet_path = root / "data" / "train-00000-of-00001.parquet"
+    fail(parquet_path.exists(), f"obsolete Viewer parquet is present: {parquet_path}", errors)
+    fail((root / "viewer_data").exists(), "obsolete viewer_data directory is present", errors)
     fail(len(list((root / "raw").rglob("*.npy"))) != expected_rows * 2, f"expected {expected_rows * 2} NPY files", errors)
     fail(len(list((root / "raw").rglob("*.npz"))) != expected_rows, f"expected {expected_rows} NPZ files", errors)
     fail(len(list((root / "comparison_sheets").rglob("*.*"))) != 93, "expected 93 comparison artifacts", errors)
@@ -97,21 +158,27 @@ def main() -> int:
         fail(len(metric_rows) != expected, f"{filename}: expected {expected} rows, found {len(metric_rows)}", errors)
         if filename in {"image_core.csv", "image_advanced.csv", "latency_components.csv", "performance_runs.csv"}:
             fail(len({row["run_id"] for row in metric_rows}) != expected_rows, f"{filename}: run_id values are not unique", errors)
-    forbidden_names = {".vendor", "advanced_metric_cache", "logs", "comfy_output", "__pycache__", ".cache", "cache"}
-    for path in root.rglob("*"):
-        fail(any(part in forbidden_names for part in path.parts), f"forbidden path: {path}", errors)
+    forbidden_names = {".vendor", "advanced_metric_cache", "logs", "comfy_output", "__pycache__", ".cache", "cache"}
+    for path in root.rglob("*"):
+        if is_local_repair_artifact(root, path):
+            continue
+        fail(any(part in forbidden_names for part in path.parts), f"forbidden path: {path}", errors)
         if path.is_file():
             fail(path.suffix.lower() in {".safetensors", ".ckpt", ".gguf", ".pyc"}, f"forbidden file: {path}", errors)
     text_suffixes = {".md", ".json", ".jsonl", ".csv", ".yaml", ".yml", ".txt", ".cff", ".py", ".ps1", ".sh"}
-    path_pattern = re.compile(r"E:\\\\Benchmark Krea 2 Turbo Formats|C:\\\\Users\\\\|GPU-[0-9a-fA-F-]{8,}|[0-9A-Fa-f]{8}:[0-9A-Fa-f]{2}:[0-9A-Fa-f]{2}\\.[0-7]|mihai", re.IGNORECASE)
-    for path in root.rglob("*"):
-        if path.is_file() and (path.suffix.lower() in text_suffixes or path.name in {"README.md", ".gitignore", ".gitattributes"}):
+    path_pattern = re.compile(r"E:\\\\Benchmark Krea 2 Turbo Formats|C:\\\\Users\\\\|GPU-[0-9a-fA-F-]{8,}|[0-9A-Fa-f]{8}:[0-9A-Fa-f]{2}:[0-9A-Fa-f]{2}\\.[0-7]|mihai", re.IGNORECASE)
+    for path in root.rglob("*"):
+        if is_local_repair_artifact(root, path):
+            continue
+        if path.is_file() and (path.suffix.lower() in text_suffixes or path.name in {"README.md", ".gitignore", ".gitattributes"}):
             if path.relative_to(root).as_posix() in {"scripts/prepare_release.py", "scripts/validate_release.py"}:
                 continue
             text = path.read_text(encoding="utf-8", errors="replace")
             fail(bool(path_pattern.search(text)), f"private machine identifier in {path.relative_to(root)}", errors)
-    link_pattern = re.compile(r"\[[^\]]*\]\(([^)]+)\)")
-    for markdown in root.rglob("*.md"):
+    link_pattern = re.compile(r"\[[^\]]*\]\(([^)]+)\)")
+    for markdown in root.rglob("*.md"):
+        if is_local_repair_artifact(root, markdown):
+            continue
         for target in link_pattern.findall(markdown.read_text(encoding="utf-8")):
             target = target.strip().strip("<>")
             if target.startswith(("http://", "https://", "mailto:", "#")):
@@ -151,28 +218,38 @@ def main() -> int:
             fail(not path.is_file(), f"checksum target missing: {relative}", errors)
             if path.is_file():
                 fail(checksum(path) != expected, f"checksum mismatch: {relative}", errors)
+        try:
+            from datasets import load_dataset
+
+            configured_splits = load_dataset(
+                "imagefolder",
+                data_files={"train": str(root / IMAGEFOLDER_GLOB)},
+                drop_labels=True,
+            )
+            fail(
+                list(configured_splits) != ["train"],
+                f"configured ImageFolder splits must be ['train'], found {list(configured_splits)!r}",
+                errors,
+            )
+            configured_dataset = configured_splits["train"]
+            fail(
+                configured_dataset.num_rows != expected_rows,
+                f"configured ImageFolder rows: {configured_dataset.num_rows}",
+                errors,
+            )
+            fail(
+                "image" not in configured_dataset.features,
+                "configured ImageFolder is missing the image feature",
+                errors,
+            )
+
+            first_image = configured_dataset[0]["image"]
+            fail(first_image.size != (1024, 1024), f"first dataset image has size {first_image.size}", errors)
+            fail(first_image.mode != "RGB", f"first dataset image has mode {first_image.mode}", errors)
+        except Exception as exc:
+            errors.append(f"dataset load failed: {exc}")
         try:
-            import pyarrow.parquet as pq
-
-            parquet_file = pq.ParquetFile(str(parquet_path))
-            fail(parquet_file.metadata.num_rows != expected_rows, f"parquet rows: {parquet_file.metadata.num_rows}", errors)
-            first_image = parquet_file.read_row_group(0, columns=["image"]).column("image")[0].as_py()
-            fail(not (isinstance(first_image, dict) and first_image.get("bytes")), "parquet image column is not embedded", errors)
-        except Exception as exc:
-            errors.append(f"parquet inspection failed: {exc}")
-        try:
-            from datasets import load_dataset
-
-            dataset = load_dataset(str(root), split="train")
-            fail(dataset.num_rows != expected_rows, f"dataset rows: {dataset.num_rows}", errors)
-            fail("image" not in dataset.features, "dataset is missing the image feature", errors)
-            _ = dataset[0]["image"]
-        except Exception as exc:
-            errors.append(f"dataset load failed: {exc}")
-        try:
-            import yaml
-
-            citation = yaml.safe_load((root / "CITATION.cff").read_text(encoding="utf-8"))
+            citation = yaml.safe_load((root / "CITATION.cff").read_text(encoding="utf-8"))
             fail(citation.get("cff-version") != "1.2.0" or citation.get("type") != "dataset", "CITATION.cff metadata is invalid", errors)
         except Exception as exc:
             errors.append(f"CITATION.cff parse failed: {exc}")