metadata
dataset_info:
features:
- name: id
dtype: string
- name: images
list:
image:
mode: RGB
- name: state
dtype: json
- name: question
dtype: json
- name: label
dtype: string
splits:
- name: train
num_bytes: 2345294519
num_examples: 4574
- name: validation
num_bytes: 265093929
num_examples: 512
- name: test_task
num_bytes: 431600231
num_examples: 833
- name: test_website
num_bytes: 306363797
num_examples: 634
- name: test_domain
num_bytes: 1268157465
num_examples: 2690
download_size: 4593730999
dataset_size: 4616509941
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test_task
path: data/test_task-*
- split: test_website
path: data/test_website-*
- split: test_domain
path: data/test_domain-*
Mind2Web decisions
Screenshot-based next-element decisions for fine-tuning Liquid d1.
This dataset accompanies the code and blog post on fine-tuning Liquid image decision models.
Dataset format
Features({
"id": Value("string"),
"images": List(Image(mode="RGB")),
"state": Json(),
"question": Json(),
"label": Value("string"),
})
Abbreviated example (image is a decoded PIL image):
{
"id": "bf94a193-a30b-45b7-a93e-4733ea6a7ed4:b88c28d6-7d12-42fc-95b3-f4267b4fd200",
"images": [image],
"state": {
"task_and_past_actions": "Find an Alaska cruise in September 2023. Dates menu opened.",
"screenshot_size_pixels": [1280, 2343],
},
"question": {
"type": "choice",
"instructions": "Select the next element to interact with.",
"criteria": {
"element_9": '{"aria_label": "October 2024", "bbox": [568, 926, 66, 50]}',
"element_17": '{"aria_label": "September 2023", "bbox": [208, 926, 66, 50]}',
},
},
"label": "element_17",
}
Load an example
Use streaming to read examples on demand without downloading the whole dataset:
from datasets import load_dataset
dataset = load_dataset(
"chrislevy/mind2web-decisions",
split="train",
streaming=True,
)
row = next(iter(dataset))
row["images"][0] # PIL image
Source and normalization
Adapted from tasksource/multimodal-mind2web, based on Multimodal-Mind2Web. The target is the correct next element. Screenshots are capped at 1 megapixel; bounding boxes use [left, top, width, height] in original-image pixels. Validation is held out by trajectory; official test splits are preserved.