metadata
language:
- en
pretty_name: ToBAC
size_categories:
- n<1K
task_categories:
- text-to-image
- image-to-text
tags:
- multimodal
- backdoor
- robustness
configs:
- config_name: default
data_files:
- split: train
path: data/*.parquet
ToBAC
Paired clean and concept-edited images accompanying Token by Token, Compromised: Backdoor Vulnerabilities in Unified Autoregressive Models, accepted at NeurIPS 2026. Code.
Format
One configuration, with 500 rows in the train split, indexed 0–499.
Each row contains a context and four images, for 2,000 images in total.
| Column | Type | Description |
|---|---|---|
context |
string | Prompt template containing {} |
clean |
Image | Clean image |
rainbow |
Image | Rainbow-flag variant |
anarchy |
Image | Anarchy-symbol variant |
pear |
Image | Pear-logo variant |
Loading
from datasets import load_dataset
dataset = load_dataset("MAI-Lab/ToBAC", split="train")
example = dataset[0]
print(example["context"])
example["clean"].show()
example["pear"].show()
Data creation
Edited images were generated from the clean image and a concept reference with
diffusers/FLUX.2-dev-bnb-4bit, revision
c30ad107542e63f222f864a8de510204394fb18a, using 50 inference steps,
guidance 4.0, and seed 42 for each image.
Intended use
ToBAC supports research on multimodal model robustness and backdoor defenses.
License
No dataset license has been specified. The code license does not automatically apply to these images.