-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathconfig.yaml
More file actions
41 lines (35 loc) · 2.03 KB
/
Copy pathconfig.yaml
File metadata and controls
41 lines (35 loc) · 2.03 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
# CrossView pipeline configuration.
# The single place that sets where raw data is read from and where the final
# benchmark is written. run.py reads this; no per-run --input is needed.
inputs:
# AgiBot & Ego-Exo4D are TWO-STEP: create_dataset reads these raw roots and
# writes a small local compiled.json (video paths + text annotations) under
# `intermediate/`, then the builder turns that into STSGs.
agibot: /nas/mars/dataset/AgiBotWorld
ego-exo4d: /nas/mars/dataset/Ego-Exo4D
# MEVA & nuScenes are ONE-STEP: the builder reads raw directly -> STSG.
# MEVA's entry point is its slot index (a catalog of the raw .yml clips); the
# raw annotations themselves live under /nas/mars/dataset/MEVA.
meva: /nas/mars/dataset/multicam_prereqs/meva/slot_index.json
# nuScenes reads pre-built per-scene scene graphs + per-instance VLM activity
# annotations under <root> (no SDK needed), plus the camera videos.
nuscenes:
root: /nas/mars/dataset/Nuscenes-MCMCQ
videos: /nas/mars/dataset/multicam_prereqs/nuscenes/videos_resized
# Models used during DATASET CONSTRUCTION (not the eval target model, which is
# passed to vqa.py --model). renderer = question phrasing + summarization
# reference writing; judge = LLM-as-judge summarization scorer.
models:
renderer: gpt-5.2
judge: gpt-5.2
# Target question counts per dataset per category. With `run.py --paper`, each
# category stops exactly at its target.
distribution:
nuscenes: {temporal: 250, event_ordering: 250, spatial: 500, counting: 500, summarization: 250}
meva: {temporal: 250, event_ordering: 250, spatial: 500, counting: 500, best_camera: 500, summarization: 250}
ego-exo4d: {temporal: 250, event_ordering: 250, best_camera: 500, summarization: 250}
agibot: {temporal: 250, event_ordering: 250, summarization: 250}
# Final benchmark output. Each dataset gets <output>/<dataset>/qa_<category>.json
output: /nas/neurosymbolic/multi-cam-dataset-final
# Local scratch for the two-step datasets' compiled.json (kept out of /nas).
intermediate: ./intermediate