khaines178/spatialai-corpora

SpatialAi corpora

Training corpora for SpatialAi, a from-scratch model that understands and edits 3D scenes. Every record is a scene, a question or instruction, the tool calls that answer it, and the answer, labelled by an exact oracle.

Folder	Stage	Train	Val	Stress
S1_frames/	S1	20,000	2,000	2,000
S2_relations/	S2	20,000	2,000	2,000
S3_frames/	S3	20,000	2,000	2,000
S4_metric/	S4	20,000	2,000	2,000
S5_structure/	S5	20,000	2,000	2,000
S6_contact/	S6	20,000	2,000	2,000
S7_transforms/	S7	20,000	2,000	2,000
S8_shapes/	S8	20,000	2,000	2,000
S9_cleanup/	S9	20,000	2,000	2,000
S9_mesh/	S9	20,000	2,000	2,000
S10_uv/	S10	20,000	2,000	2,000
S11_ops/	S11	20,000	2,000	2,000
S12_rigs/	S12	20,000	2,000	2,000
S13_kinematics/	S13	20,000	2,000	2,000
S14_compose/	S14	20,000	2,000	2,000
LT2_dialog/	LT2	20,000	2,000	2,000
V2_edits/	V2	20,000	2,000	2,000
V3_create/	V3	20,000	2,000	2,000
V4_bevel/	V4	20,000	2,000	2,000
V5_inset/	V5	20,000	2,000	2,000
V6_extrude/	V6	20,000	2,000	2,000
V7_loopcut/	V7	20,000	2,000	2,000

Each folder holds train.jsonl, val.jsonl (held-out wordings and placements), stress.jsonl (turned, mirrored, restated and crowded scenes), manifest.json (seed, counts, settings) and answers.json. mixes/ holds the weights that combine corpora into one run.

Each build is tagged, so a run can pull exactly the data it trained on. Builds: ladder-v1, ladder-v2, ladder-v3, ladder-v4, ladder-v5, ladder-v6, ladder-v7, ladder-v8, ladder-v9.

from huggingface_hub import snapshot_download
snapshot_download("khaines178/spatialai-corpora", repo_type="dataset", revision="ladder-v9",
                  allow_patterns=["S1_frames/*"], local_dir="data")