LSI-108K

Interaction-derived supervision for spatial reasoning

LSI-108K is the dataset introduced in Spatial-Interactor: Learning Spatial Reasoning through Interaction with the Observable Physical World. It contains 107,518 verifiable QA pairs organized as a three-level spatial interaction curriculum.

Overview

Presentation

A three-level curriculum

Split	Samples	Curriculum role
L1	15,109	Passive world-state transitions
L2	69,487	Active self-state transitions
L3	22,922	Long-horizon transition integration
Total	107,518	Interaction-derived spatial QA

L1 follows object and robot state changes under a fixed observer. L2 follows camera motion and its spatial consequences. L3 integrates consecutive camera transitions into paths, turns, revisits, and reverse routes.

From trajectories to QA

Simulator actions, scene states, camera poses, object tracks, and robot trajectories provide geometric ground truth before task templates produce QA pairs. This keeps supervision tied to observable physical change.

Load

from datasets import load_dataset

dataset = load_dataset("kagakouko/LSI-108K")
print(dataset)

Each row contains an id, a ShareGPT-style conversations list, media_paths, media_type (image or video), and metadata with the curriculum level, task type, source identifier, and geometry-derived ground truth. Structured geometry is stored as compact JSON in metadata.gt_json. The used_in_reported_sft field marks the 82,596 L1-L2 records used by the reported SFT stage.

For LLaMA-Factory-style training, assign media_paths to images when media_type == "image", and to videos otherwise.

Media and training protocol

This repository releases annotations and portable relative media references; third-party images and videos remain governed by their original licenses and are not redistributed. Obtain the source assets from their original providers and map them under a local media root.

The reported SFT run uses 13,109 single-transition L1 records and all 69,487 L2 records together with the public spatial QA mixture described in the paper. The remaining 2,000 multi-stage operation records belong to L1 but were not included in that run. OPD uses the selected L3 training split together with long-horizon VSTI records and question-independent privileged transition traces.

Construction and validation code is available in the Spatial-Interactor repository.

Citation

For citation, use the project BibTeX.

License

The annotations are provided for research use. Source media are governed by their original licenses and are not included here. Users are responsible for complying with the terms of every upstream source they obtain.