Update dataset card: train/test now use matching clean modality names

README.md

@@ -27,32 +27,15 @@ K600-MM is the multimodal video dataset used to pretrain and evaluate [A2A-Video
 ## Directory structure
 
 ```
-train/<modality_folder>/vol_00.tar ... vol_63.tar   # 64 shards per modality
-test/<modality_folder>/vol_00.tar  ... vol_63.tar   # 64 shards per modality
+train/<modality>/vol_00.tar ... vol_63.tar   # 64 shards per modality
+test/<modality>/vol_00.tar  ... vol_63.tar   # 64 shards per modality
 ```
 
-Each `vol_XX.tar` is a [WebDataset](https://github.com/webdataset/webdataset) shard: many individually-named files, one (or a few) per video clip, sharing a filename stem across modalities so samples can be aligned by name. `train` and `test` use the same layout, but **different modality folder names** -- see below.
+Each `vol_XX.tar` is a [WebDataset](https://github.com/webdataset/webdataset) shard: many individually-named files, one (or a few) per video clip, sharing a filename stem across modalities so samples can be aligned by name.
 
-### Modality folder names
+`train/` and `test/` use matching modality folder names -- both use A2A-Video's internal modality names directly, so no name-mapping config is needed to consume this dataset as downloaded:
 
-`train/` folder names carry legacy processing-pipeline naming and need to be mapped to A2A-Video's internal modality names via `modality_name_map` in the data config (see [an example config](https://github.com/EPFL-VILAB/A2A-Video/blob/main/cfgs/default/4m/data/cc12m/main/4m_b_video_all_modalities_final_config.yaml)):
-
-| `train/` folder | Internal modality name |
-|---|---|
-| `rgb_video_tokens_128_23rd_sept` | `tok_video_rgb@128` |
-| `normals_npz_pseudolabels_video_tokens_128_23rd_sept` | `tok_video_normal@128` |
-| `depth_npz_pseudolabels_video_tokens_128_23rd_sept` | `tok_video_depth@128` |
-| `opticalflow_video_tokens_128_23rd_sept` | `tok_video_opticalflow@128` |
-| `siglip2_video_tokens_128` | `tok_video_siglipv2@224` |
-| `DINOv2-B14_video_tokens_128` | `tok_video_dinov2@224` |
-| `rgb_video_tokens_256_vjepa_tar` | `tok_video_vjepa@224` |
-| `detection` | `det` |
-| `saved_human_poses_pseudolabeled_with_tokenized_poses` | `human_poses` |
-| `caption`, `class_condition`, `transcription`, `crop_settings` | same name, no mapping needed |
-
-`test/` folder names already match A2A-Video's internal modality names directly (`tok_video_rgb@128`, `tok_video_depth@128`, `tok_video_dinov2@224`, `tok_video_flextok@128`, `tok_video_normal@128`, `tok_video_opticalflow@128`, `tok_video_siglipv2@224`, `tok_video_vjepa@224`, `det`, `human_poses`, `caption`, `class_condition`, `transcription`, `qwenvl_caption`, `crop_settings`) -- no `modality_name_map` entry needed for these.
-
-Both splits also include a plain `rgb` folder (raw/untokenized video) not required for standard pretraining -- it's there for tokenizer training and visualization use cases.
+`caption`, `class_condition`, `crop_settings`, `det`, `human_poses`, `rgb`, `tok_video_depth@128`, `tok_video_dinov2@224`, `tok_video_normal@128`, `tok_video_opticalflow@128`, `tok_video_rgb@128`, `tok_video_siglipv2@224`, `tok_video_vjepa@224`, `transcription` (14 folders -- the released model's 12 `in_domains`/`out_domains` modalities, plus `crop_settings` and raw, untokenized `rgb`).
 
 ## Downloading
 
@@ -67,7 +50,7 @@ from huggingface_hub import snapshot_download
 
 snapshot_download(
     repo_id="EPFL-VILAB/K600-MM", repo_type="dataset", local_dir="/path/to/k600-mm",
-    allow_patterns=["train/caption/*", "train/rgb_video_tokens_128_23rd_sept/*"],
+    allow_patterns=["train/caption/*", "train/tok_video_rgb@128/*"],
 )
 ```
 
@@ -76,7 +59,7 @@ snapshot_download(
 Point `data_path` in your data config at wherever this repo ended up on disk, following the bracket-notation shown in [the example data config](https://github.com/EPFL-VILAB/A2A-Video/blob/main/cfgs/default/4m/data/cc12m/main/4m_b_video_all_modalities_final_config.yaml):
 
 ```yaml
-data_path: '/path/to/k600-mm/train/[crop_settings,caption,rgb_video_tokens_128_23rd_sept,...]/vol_{00..63}.tar'
+data_path: '/path/to/k600-mm/train/[crop_settings,caption,tok_video_rgb@128,...]/vol_{00..63}.tar'
 ```
 
 ## License