Rewrite dataset card and keep one default config

README.md

@@ -1,5 +1,5 @@
 ---
-pretty_name: Diffusion4D Animated Raw Assets
+pretty_name: Diffusion4D Animated Assets
 language:
 - en
 tags:
@@ -16,68 +16,71 @@ configs:
   data_files:
   - split: train
     path: data/train-*.parquet
-- config_name: preview_sample
-  data_files:
-  - split: train
-    path: preview_sample/train-*.parquet
-- config_name: with_previews
-  data_files:
-  - split: train
-    path: with_previews/train-*.parquet
-- config_name: video_sample
-  data_files:
-  - split: train
-    path: video_sample/train-*.parquet
-- config_name: with_media
-  data_files:
-  - split: train
-    path: with_media/train-*.parquet
 ---
 
-# Diffusion4D Animated Raw Assets
+# Diffusion4D Animated Assets
+
+This dataset provides animated 3D assets referenced by Diffusion4D and
+Objaverse-XL in a directly browsable format. The default split contains 67,988
+rows. Each row includes metadata, a preview image, and a short preview video so
+that assets can be inspected in the Hugging Face Data Studio without first
+downloading the original 3D file.
+
+The repository also mirrors available raw assets and keeps their original
+source links and hashes. The current release is intentionally broad and may
+contain broken, static, duplicated, or semantically irrelevant assets.
+
+## Curation In Progress
+
+Our goal is to identify high-quality skeletal animations suitable for training
+and evaluating motion and 3D understanding models. We are currently annotating
+every asset with Qwen3-VL-30B-A3B-Instruct using four frames sampled from its
+animation.
+
+The automatic annotations include:
+
+- a broad semantic class and a more specific object class;
+- a short object description;
+- a description of the visible motion;
+- whether meaningful animation is visible and its strength;
+- visual quality, visible defects, and model confidence.
 
-An evolving, metadata-first mirror of animated 3D assets referenced by the
-Diffusion4D / Objaverse-XL metadata. This initial release is intentionally
-unfiltered: it preserves assets that may be low quality, semantically
-irrelevant, technically invalid, duplicated by appearance, or difficult to
-import.
+These labels are machine generated and should not be treated as human-verified
+ground truth. Failed or malformed model responses will be processed again in a
+separate repair pass.
 
-## Layout
+## Planned Animated Subset
 
-- `assets/{format}/{asset_id}/{filename}`: raw source assets.
-- `previews/glb/{preview_id}/preview.mp4`: available animation previews.
-- `data/train-*.parquet`: stable, sharded Dataset Viewer table.
-- `metadata/assets.parquet`: metadata snapshot used by the release tooling.
-- `metadata/assets.jsonl`: lossless JSONL mirror of the table.
-- `metadata/summary.json`: release coverage counts.
+We plan to publish a separate curated subset of animated animals, humanoids,
+robots, and other articulated characters. Selection will combine technical
+metadata, visual inspection, and Qwen annotations.
 
-## Metadata
+For this subset we also plan to provide:
 
-Records include source URI and SHA-256, source format, download status,
-skeleton and animation statistics, skinning and texture flags, technical QC,
-preview paths, and placeholders for later VLM annotations.
+- renders from multiple camera angles;
+- per-frame DINO visual features;
+- data arranged for training models such as MocapAnything v2.
 
-`technical_pass` currently requires at least 20 bones, at least 20 animation
-frames, skinning weights, and a detected image texture. It is a technical flag,
-not a semantic quality label.
+## Data Fields
 
-## Release Status
+The table includes identifiers and source metadata, file format and hashes,
+animation and skeleton statistics when available, technical inspection flags,
+and embedded preview media. Additional Qwen annotation columns will be added
+after the full annotation and validation passes finish.
 
-This repository is uploaded incrementally. A row can exist before its raw asset
-or preview has been mirrored. Check `asset_available`, `preview_available`, and
-`annotation_status` before use.
+Use availability and quality fields when filtering. A row being present does
+not mean that its raw asset is valid, complete, or suitable for training.
 
-## Provenance And Licenses
+## Provenance And Licensing
 
-The records originate from Diffusion4D / Objaverse-XL metadata and preserve the
-original source URI. Assets come from many upstream repositories and do not
-share one normalized license. `license_status=upstream_not_normalized` means the
-downstream user must inspect the linked upstream repository and comply with its
-license before redistribution or commercial use. This repository does not
-relicense third-party assets.
+The records and official renders originate from Diffusion4D and Objaverse-XL.
+Raw assets come from many upstream repositories and do not share a single
+normalized license. Source URLs are retained where available. Users are
+responsible for checking the license of each upstream asset before
+redistribution or commercial use. This dataset does not relicense third-party
+content.
 
-## Planned Curation
+## Acknowledgements
 
-Later revisions will add four-frame contact sheets, Qwen-VL object and motion
-descriptions, semantic classes, visual defect labels, normalized licenses,
-duplicate groups, and curated training subsets while retaining this raw index.
+Please cite the original Diffusion4D and Objaverse-XL projects when using this
+dataset. This repository is an independent packaging and curation effort.