specialsuzuka/SeekVLN-Dataset

SeekVLN Dataset

The 4K training dataset for Seek Before You Move: Evidence Seeking for Progress Grounding in Vision-Language Navigation.

Paper: https://arxiv.org/abs/2609.37353

SFT checkpoint: https://huggingface.co/specialsuzuka/SeekVLN-FRG-SFT

This repository is being prepared for release and remains private pending licensing and release review. No new license is granted by this draft card; underlying data rights and redistribution conditions must be reviewed before public release.

Contents

• annotations/annotations.json: original decision-level annotations, including instructions, navigation history, seek-view references, and responses.
• annotations/seekvln_sft_4k.json: image references and multi-turn SFT conversations.
• images/train-*.tar: real image files, grouped by episode; no symbolic links. Only images referenced by the released annotations or SFT conversations are included.
• stats.json: exact counts and shard list.
• SHA256SUMS: integrity hashes.

The training set contains 4,162 episodes and 111,141 decision-level records: 68,795 NAV and 42,346 SEEK records. These are not 111,141 independent episodes. No validation or test split is included.

Download and Layout

from pathlib import Path
import tarfile
from huggingface_hub import snapshot_download

root = Path(snapshot_download("specialsuzuka/SeekVLN-Dataset", repo_type="dataset"))
for shard in sorted((root / "images").glob("train-*.tar")):
    with tarfile.open(shard) as archive:
        archive.extractall(root, filter="data")  # Python 3.12+

After extraction, resolve every annotation or SFT image path relative to R2R/train/. For example, 1/frame_0.jpg resolves to R2R/train/1/frame_0.jpg. JSON files preserve their original schemas and relative image paths. The TAR shards are storage bundles, not a WebDataset sample schema; use the JSON files to reconstruct trajectory history and conversations. A private repository requires an authorized Hugging Face login.

The policy uses <nav> for direct navigation and <seek> to request additional observations before progress reasoning and navigation. Some original annotations also reference a back view; these references and images are preserved, while the SFT conversations use their original selected views.

Annotation Configurations

The decisions and sft configurations each contain a train split. They expose the two distinct JSON schemas separately; image fields remain relative-path strings, not embedded images. Download and extract the image shards when consuming image content.

from datasets import load_dataset

decisions = load_dataset("specialsuzuka/SeekVLN-Dataset", "decisions", split="train", token=True)
sft = load_dataset("specialsuzuka/SeekVLN-Dataset", "sft", split="train", token=True)