Navinvue commited on
Commit
fe72315
·
verified ·
1 Parent(s): 64ea2f5

Add dataset card

Browse files
Files changed (1) hide show
  1. README.md +71 -0
README.md ADDED
@@ -0,0 +1,71 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ language:
4
+ - en
5
+ pretty_name: HomeSafeBench
6
+ tags:
7
+ - embodied-ai
8
+ - vision-language
9
+ - home-safety
10
+ - virtualhome
11
+ configs:
12
+ - config_name: benchmark
13
+ default: true
14
+ data_files:
15
+ - split: train
16
+ path: data/train/*.json
17
+ - split: test
18
+ path: data/test/*.json
19
+ - config_name: sft_action
20
+ data_files:
21
+ - split: train
22
+ path: cueback/sft_action.jsonl
23
+ - config_name: sft_both
24
+ data_files:
25
+ - split: train
26
+ path: cueback/sft_both.jsonl
27
+ - config_name: cueback
28
+ data_files:
29
+ - split: train
30
+ path: cueback/cueback.jsonl
31
+ ---
32
+
33
+ # HomeSafeBench
34
+
35
+ This repository contains the official data release for the paper
36
+ [HomeSafeBench: Benchmarking Embodied Vision-Language Models in Free-Exploration Home Safety Inspection](https://arxiv.org/abs/2509.23690).
37
+
38
+ HomeSafeBench is a benchmark for free-exploration home safety inspection with
39
+ embodied vision-language models, built on
40
+ [VirtualHome](https://github.com/xavierpuigf/virtualhome).
41
+
42
+ ## Dataset Contents
43
+
44
+ - `data/train/`: 3,400 training tasks.
45
+ - `data/test/`: 1,000 human-validated evaluation tasks.
46
+ - `cueback/sft_action.jsonl`: executable action trajectories without reasoning.
47
+ - `cueback/sft_both.jsonl`: trajectories with directly generated reasoning.
48
+ - `cueback/cueback.jsonl`: trajectories with CueBack-constructed reasoning.
49
+ - `cueback/images/`: referenced first-person observations.
50
+
51
+ The benchmark format, setup instructions, and evaluation code are documented
52
+ in the [HomeSafeBench code repository](https://github.com/BITHLP/HomeSafeBench).
53
+
54
+ ## Citation
55
+
56
+ ```bibtex
57
+ @misc{yao2026homesafebenchbenchmarkembodiedvisionlanguage,
58
+ title={HomeSafeBench: Benchmarking Embodied Vision-Language Models in Free-Exploration Home Safety Inspection},
59
+ author={Jiashu Yao and Haoyu Wen and Siyuan Gao and Yuhang Guo and Zeming Liu and Heyan Huang},
60
+ year={2026},
61
+ eprint={2509.23690},
62
+ archivePrefix={arXiv},
63
+ primaryClass={cs.CV},
64
+ url={https://arxiv.org/abs/2509.23690},
65
+ }
66
+ ```
67
+
68
+ ## License
69
+
70
+ The HomeSafeBench dataset is released under the
71
+ [Creative Commons Attribution 4.0 International License](https://creativecommons.org/licenses/by/4.0/).