Dataset Viewer
Auto-converted to Parquet Duplicate
scene_index
int32
0
3.29k
video_filename
stringlengths
15
15
video
video
5.12
5.12
question_id
int32
11
15
question
stringlengths
34
79
question_type
stringclasses
1 value
question_subtype
stringclasses
0 values
program
listlengths
7
12
answer
stringclasses
0 values
choices
listlengths
1
4
annotation_json
large_stringlengths
82.9k
162k
0
video_00000.mp4
13
Which event will happen if the cylinder is removed?
counterfactual
null
[ "all_events", "objects", "cylinder", "filter_shape", "unique", "get_counterfact", "belong_to" ]
null
[ { "choice_id": 0, "choice": "The blue rubber sphere collides with the cube", "program": [ "all_events", "objects", "blue", "filter_color", "rubber", "filter_material", "sphere", "filter_shape", "unique", "filter_collision", "objects", ...
{"scene_index": 0, "video_filename": "video_00000.mp4", "object_property": [{"object_id": 0, "color": "blue", "material": "rubber", "shape": "sphere"}, {"object_id": 1, "color": "gray", "material": "rubber", "shape": "sphere"}, {"object_id": 2, "color": "cyan", "material": "metal", "shape": "cube"}, {"object_id": 3, "c...
0
video_00000.mp4
14
Without the metal sphere, which event will not happen?
counterfactual
null
["all_events","objects","metal","filter_material","sphere","filter_shape","unique","get_counterfact"(...TRUNCATED)
null
[{"choice_id":0,"choice":"The blue rubber sphere and the gray sphere collide","program":["all_events(...TRUNCATED)
"{\"scene_index\": 0, \"video_filename\": \"video_00000.mp4\", \"object_property\": [{\"object_id\":(...TRUNCATED)
1
video_00001.mp4
14
Without the cube, which event will happen?
counterfactual
null
[ "all_events", "objects", "cube", "filter_shape", "unique", "get_counterfact", "belong_to" ]
null
[{"choice_id":0,"choice":"The cylinder collides with the gray object","program":["all_events","objec(...TRUNCATED)
"{\"scene_index\": 1, \"video_filename\": \"video_00001.mp4\", \"object_property\": [{\"object_id\":(...TRUNCATED)
1
video_00001.mp4
15
What will not happen without the cylinder?
counterfactual
null
[ "all_events", "objects", "cylinder", "filter_shape", "unique", "get_counterfact", "belong_to", "negate" ]
null
[{"choice_id":0,"choice":"The cube and the purple sphere collide","program":["all_events","objects",(...TRUNCATED)
"{\"scene_index\": 1, \"video_filename\": \"video_00001.mp4\", \"object_property\": [{\"object_id\":(...TRUNCATED)
2
video_00002.mp4
13
Which event will happen if the cyan cube is removed?
counterfactual
null
["all_events","objects","cyan","filter_color","cube","filter_shape","unique","get_counterfact","belo(...TRUNCATED)
null
[{"choice_id":0,"choice":"The red cube collides with the gray sphere","program":["all_events","objec(...TRUNCATED)
"{\"scene_index\": 2, \"video_filename\": \"video_00002.mp4\", \"object_property\": [{\"object_id\":(...TRUNCATED)
3
video_00003.mp4
14
Which event will happen without the yellow cylinder?
counterfactual
null
["all_events","objects","yellow","filter_color","cylinder","filter_shape","unique","get_counterfact"(...TRUNCATED)
null
[{"choice_id":0,"choice":"The rubber cylinder and the sphere collide","program":["all_events","objec(...TRUNCATED)
"{\"scene_index\": 3, \"video_filename\": \"video_00003.mp4\", \"object_property\": [{\"object_id\":(...TRUNCATED)
4
video_00004.mp4
13
What will not happen if the sphere is removed?
counterfactual
null
[ "all_events", "objects", "sphere", "filter_shape", "unique", "get_counterfact", "belong_to", "negate" ]
null
[{"choice_id":0,"choice":"The metal cube and the cyan cylinder collide","program":["all_events","obj(...TRUNCATED)
"{\"scene_index\": 4, \"video_filename\": \"video_00004.mp4\", \"object_property\": [{\"object_id\":(...TRUNCATED)
4
video_00004.mp4
14
Which of the following will happen if the blue object is removed?
counterfactual
null
[ "all_events", "objects", "blue", "filter_color", "unique", "get_counterfact", "belong_to" ]
null
[{"choice_id":0,"choice":"The metal cube collides with the cyan cylinder","program":["all_events","o(...TRUNCATED)
"{\"scene_index\": 4, \"video_filename\": \"video_00004.mp4\", \"object_property\": [{\"object_id\":(...TRUNCATED)
5
video_00005.mp4
14
What will not happen without the blue sphere?
counterfactual
null
["all_events","objects","blue","filter_color","sphere","filter_shape","unique","get_counterfact","be(...TRUNCATED)
null
[{"choice_id":0,"choice":"The cyan cylinder and the gray cylinder collide","program":["all_events","(...TRUNCATED)
"{\"scene_index\": 5, \"video_filename\": \"video_00005.mp4\", \"object_property\": [{\"object_id\":(...TRUNCATED)
6
video_00006.mp4
14
Without the red sphere, what will happen?
counterfactual
null
["all_events","objects","red","filter_color","sphere","filter_shape","unique","get_counterfact","bel(...TRUNCATED)
null
[{"choice_id":0,"choice":"The blue cube collides with the purple object","program":["all_events","ob(...TRUNCATED)
"{\"scene_index\": 6, \"video_filename\": \"video_00006.mp4\", \"object_property\": [{\"object_id\":(...TRUNCATED)
End of preview. Expand in Data Studio

CLEVRER

A Hugging Face packaging of CLEVRER. The question types are the Hugging Face configs: descriptive, explanatory, predictive, and counterfactual. Each has the original train/validation/test split boundaries, with one row per question of that type. Question fields are top-level columns, not a nested questions list. Each question row embeds the matching video's actual MP4 bytes. Empty splits/configs are omitted, which can happen when packaging a small subset.

The original release has 10,000 train, 5,000 validation and 5,000 test videos. Consult the config/split sizes above for the number of question rows.

Each row contains one original MP4 embedded as bytes in Parquet, not an external link, together with:

  • scene_index and video_filename: original scene identifier and filename.
  • video: Hugging Face Video(decode=False); video["bytes"] contains the MP4.
  • question_id, question, question_type, question_subtype, program, answer, and choices: top-level question columns. choices is a list of options, with each option's ID, text, program and answer. Multiple choices can be correct, or none can be correct. Missing fields are null; descriptive questions have an empty choices list. question_id is local to a scene; (scene_index, question_id) identifies a question.
  • annotation_json: lossless scene JSON (object properties, trajectories, collisions). Null for test, or if annotations were omitted during packaging. Use json.loads(row["annotation_json"]) to read an available annotation.

Test answers and programs are not provided by the original release and remain null here. The separate model-derived object-mask/proposal archive is not included. No videos are transcoded and no train/validation/test scenes are mixed.

from datasets import load_dataset, Video

ds = load_dataset("ghomasHudson/clevrer", "descriptive")  # Select question_type config.
row = ds["train"][0]  # Or any available split.
with open(row["video_filename"], "wb") as f:
    f.write(row["video"]["bytes"])
print(row["question"], row["answer"], row["choices"])
# Optional: install compatible torch/torchcodec and FFmpeg to decode frames.
decoded = ds["train"].cast_column("video", Video(decode=True))

Source and license

Videos, annotations and questions are downloaded from the links on the official dataset page, using HTTPS. The upstream README specifies CC0. This is a repackaging, not a new dataset or an official author-maintained mirror.

Citation

@inproceedings{CLEVRER2020ICLR,
  author = {Kexin Yi and Chuang Gan and Yunzhu Li and Pushmeet Kohli and
            Jiajun Wu and Antonio Torralba and Joshua B. Tenenbaum},
  title = {{CLEVRER:} Collision Events for Video Representation and Reasoning},
  booktitle = {ICLR},
  year = {2020}
}
Downloads last month
284