Dataset Viewer
Auto-converted to Parquet Duplicate
text
stringlengths
5
42
77371
10002158834_0 on 10002158834_1
10002158834_10 above 10002158834_1
10002158834_2 on 10002158834_0
10002158834_3 above 10002158834_0
10002158834_3 above 10002158834_1
10002158834_3 behind 10002158834_0
10002158834_4 on 10002158834_5
10002158834_6 behind 10002158834_0
10002158834_6 has 10002158834_7
10002158834_6 on 10002158834_5
10002158834_8 at 10002158834_5
10002158834_9 above 10002158834_1
10002158834_9 on 10002158834_8
10003464684_0 has 10003464684_4
10003464684_0 has 10003464684_5
10003464684_0 has 10003464684_9
10003464684_1 has 10003464684_2
10003464684_10 wearing 10003464684_11
10003464684_12 behind 10003464684_0
10003464684_14 wearing 10003464684_15
10003464684_16 near 10003464684_1
10003464684_17 behind 10003464684_0
10003464684_17 wearing 10003464684_18
10003464684_20 above 10003464684_32
10003464684_20 has 10003464684_31
10003464684_20 near 10003464684_20
10003464684_20 on 10003464684_21
10003464684_22 wearing 10003464684_23
10003464684_26 behind 10003464684_25
10003464684_27 on 10003464684_28
10003464684_29 near 10003464684_0
10003464684_3 on 10003464684_1
10003464684_30 near 10003464684_0
10003464684_33 above 10003464684_34
10003464684_6 has 10003464684_7
10003464684_6 wearing 10003464684_24
10003464684_8 over 10003464684_1
10003466765_0 on 10003466765_10
10003466765_1 has 10003466765_2
10003466765_11 above 10003466765_1
10003466765_11 above 10003466765_15
10003466765_11 on 10003466765_12
10003466765_16 on 10003466765_17
10003466765_3 near 10003466765_1
10003466765_5 has 10003466765_6
10003466765_5 in 10003466765_4
10003466765_5 under 10003466765_3
10003466765_7 on 10003466765_1
10003466765_9 holding 10003466765_1
10004827943_0 in front of 10004827943_13
10004827943_0 in front of 10004827943_2
10004827943_0 left of 10004827943_14
10004827943_0 on 10004827943_5
10004827943_0 under 10004827943_12
10004827943_1 above 10004827943_2
10004827943_1 above 10004827943_6
10004827943_10 on 10004827943_0
10004827943_11 near 10004827943_1
10004827943_11 under 10004827943_1
10004827943_12 above 10004827943_0
10004827943_13 behind 10004827943_14
10004827943_15 in 10004827943_16
10004827943_2 behind 10004827943_0
10004827943_2 near 10004827943_0
10004827943_2 on 10004827943_5
10004827943_2 under 10004827943_1
10004827943_3 near 10004827943_14
10004827943_4 in front of 10004827943_2
10004827943_4 standing on 10004827943_7
10004827943_6 behind 10004827943_2
10004827943_9 on 10004827943_0
10007531244_0 has 10007531244_4
10007531244_0 on 10007531244_1
10007531244_11 on 10007531244_10
10007531244_12 hanging from 10007531244_13
10007531244_2 hanging from 10007531244_14
10007531244_3 against 10007531244_1
10007531244_3 lying on 10007531244_1
10007531244_3 near 10007531244_0
10007531244_3 near 10007531244_2
10007531244_3 under 10007531244_0
10007531244_5 near 10007531244_1
10007531244_5 on 10007531244_9
10007531244_5 under 10007531244_0
10007531244_6 on 10007531244_7
10007531244_6 under 10007531244_8
10008170616_0 wearing 10008170616_1
10008170616_0 wearing 10008170616_5
10008170616_11 near 10008170616_18
10008170616_13 near 10008170616_7
10008170616_16 over 10008170616_13
10008170616_17 on 10008170616_8
10008170616_19 wearing 10008170616_19
10008170616_2 wearing 10008170616_6
10008170616_3 holding 10008170616_12
10008170616_3 wearing 10008170616_4
10008170616_8 has 10008170616_10
10008170616_8 has 10008170616_15
10008170616_8 holding 10008170616_9
End of preview. Expand in Data Studio

Spatial-VRD-Trans (Geometry-MMKG, data v0)

Transductive control for callmezuy/spatial-vrd: the same 107,404 triples, entities and features, re-split so that every valid / test entity occurs in train. Comparing a model on both repos separates the effect of the inductive setting from the effect of the data. Code: github.com/callmezuy/Geometry-MMKG (scripts/make_transductive_split.py --src data/Spatial-VRD --dst data/Spatial-VRD-Trans --stratify relation).

Split Triples Entities
train 77,371 83,552
valid 8,578 14,210
test 21,455 27,937

93,809 entities in entity2id.txt (identical to spatial-vrd), 83,552 of them in triples; 52 relations, all in every split. The hold-out is stratified by relation (valid 8% / test 20%), and a triple is held out only if both of its entities keep a triple in train. split_stats.json has the details.

Files

Same formats as spatial-vrd (see its card): split files with a count line, OpenKE *2id.txt, entity2text.txt, relation_category.tsv, and the same features (ent_rgb_emb.pt, ent_depth_emb.pt, ent_text_emb.pt, ent_rgb_depth_mean_emb.pt, ent_rgb_depth_concat_emb.pt, ent_img_mask.pt, ent_fill.pt, ent_depth_stats.pt [93809, 9] with depth_stats.json). Only the fill differs: the 2 entities without a usable box get the mean feature (and depth statistics) of the train entities of this split.

Versions

v0 (18/11/2026) is the first version of this repo (earlier repos with this name were deleted). Pin it with snapshot_download("callmezuy/spatial-vrd-trans", repo_type="dataset", revision="v0"). On 19/11/2026 ent_depth_stats.pt and depth_stats.json were added and the tag moved; no other file changed.

License

Derived from the VRD dataset (Lu et al., 2016); use under its terms.

Downloads last month
128