ProSpace-Train / README.md
shamanneo's picture
Upload README.md with huggingface_hub
0552e5c verified
|
Raw History Blame Contribute Delete
970 Bytes
metadata
task_categories:
  - video-text-to-text
language:
  - en
tags:
  - spatial-reasoning
  - streaming-video
  - proactive
size_categories:
  - 100K<n<1M
configs:
  - config_name: 100k
    data_files: prospace_train_100k.jsonl.gz
  - config_name: 400k
    data_files: prospace_train_400k.jsonl.gz

ProSpace-Train

Training streams of Thinking in Streams: Proactive Spatial Reasoning with Vision-Language Models, built from the official training splits of ScanNet++, ScanNet, CA-1M and Matterport3D.

File Content
prospace_train_100k.jsonl.gz balanced 100K set (the set of the paper's fine-tuned rows)
prospace_train_400k.jsonl.gz extended 400K set
val_es.jsonl, val_poll.jsonl validation streams used for checkpoint selection

The frames are not redistributed. Trainers (native streaming and prefix polling): https://github.com/shamanneo/prospace

hf download shamanneo/ProSpace-Train --repo-type dataset --local-dir data/train