Datasetpublic

syp115/StreamVLN-Trajectory-Data-ZIP

StreamVLN R2R + RxR — one ZIP per trajectory A storage-only repack of cywan/StreamVLN-Trajectory-Data, containing the R2R and RxR RGB training trajectories used by our Qwen3-VL navigation baseline. Credit for the images and annotations belongs to the original dataset authors. The upstream license notice specifies CC BY-NC-SA 4.0; see the upstream dataset for its full terms. Split Trajectory ZIPs Original JPEG frames R2R 10,819 647,622 RxR 19,990 1,901,165 Total 30… See the full description on the dataset page: https://huggingface.co/datasets/syp115/StreamVLN-Trajectory-Data-ZIP.

sourceHugging Facecc-by-nc-sa-4.0updated 26d agoView on Hugging Face

1likes383downloads

Dataset Card

StreamVLN R2R + RxR — one ZIP per trajectory

A storage-only repack of cywan/StreamVLN-Trajectory-Data, containing the R2R and RxR RGB training trajectories used by our Qwen3-VL navigation baseline. Credit for the images and annotations belongs to the original dataset authors. The upstream license notice specifies CC BY-NC-SA 4.0; see the upstream dataset for its full terms.

Split	Trajectory ZIPs	Original JPEG frames
R2R	10,819	647,622
RxR	19,990	1,901,165
Total	30,809	2,548,787

Each ZIP uses ZIP_STORED, preserving original JPEG bytes without recompression. Each archive contains rgb/000.jpg, rgb/001.jpg, etc. Read selected frames directly from the ZIP; do not extract the dataset into millions of individual files.

Layout

text

R2R/annotations_v1-3.json
R2R/images/{scene}/{trajectory}.zip
RxR/annotations.json
RxR/images/{scene}/{trajectory}.zip

The scene subdirectories keep each directory below Hugging Face's 10,000-entry limit. Only the annotation video path is updated to images/{scene}/{trajectory}. All other annotation fields, including instructions, actions, and trajectory order, are unchanged. All source frame bytes were verified during packing.

There are 30,809 ZIP files, two annotation files, this README, and the Hub-generated .gitattributes: 30,813 repository files in total after upload. Upload caches and local verification files are excluded.

Download and train

Download the repository with hf download syp115/StreamVLN-Trajectory-Data-ZIP --repo-type dataset --local-dir /your/data/StreamVLN-Trajectory-Data-ZIP.

In the ZIP-aware Qwen3-VL VLN baseline, set annotation_path to the annotation JSON and image_root to the corresponding R2R/images or RxR/images directory. The existing dataset reader resolves the scene-prefixed paths automatically. History sampling and action supervision are unchanged.

This repack does not add depth supervision, inverse-dynamics tasks, model weights, or training results.