Datasets:
boyuanzheng010
commited on
Commit
•
f27b636
1
Parent(s):
1e98bc5
Update README.md
Browse files
README.md
CHANGED
@@ -59,7 +59,9 @@ dataset_info:
|
|
59 |
|
60 |
### Dataset Summary
|
61 |
|
62 |
-
Multimodal-Mind2Web is the multimodal version of [Mind2Web](https://osu-nlp-group.github.io/Mind2Web/), a dataset for developing and evaluating generalist agents
|
|
|
|
|
63 |
|
64 |
## Dataset Structure
|
65 |
|
@@ -69,19 +71,18 @@ Multimodal-Mind2Web is the multimodal version of [Mind2Web](https://osu-nlp-grou
|
|
69 |
- test_website: 1019 actions from 142 tasks. Websites are not seen during training.
|
70 |
- test_domain: 4060 actions from 694 tasks. Entire domains are not seen during training.
|
71 |
|
72 |
-
The three
|
73 |
|
74 |
-
Each line in the dataset is an action consisting of screenshot image, HTML text and other fields required for action prediction, for the convenience of inference.
|
75 |
|
76 |
### Data Fields
|
77 |
-
|
78 |
- "annotation_id" (str): unique id for each task
|
79 |
- "website" (str): website name
|
80 |
- "domain" (str): website domain
|
81 |
- "subdomain" (str): website subdomain
|
82 |
- "confirmed_task" (str): task description
|
83 |
- "action_reprs" (list[str]): human readable string representation of the action sequence
|
84 |
-
- "screenshot" (str): path to the webpage screenshot image corresponding to the HTML
|
85 |
- "action_uid" (str): unique id for each action (step)
|
86 |
- "raw_html" (str): raw html of the page before the action is performed
|
87 |
- "cleaned_html" (str): cleaned html of the page before the action is performed
|
|
|
59 |
|
60 |
### Dataset Summary
|
61 |
|
62 |
+
Multimodal-Mind2Web is the multimodal version of [Mind2Web](https://osu-nlp-group.github.io/Mind2Web/), a dataset for developing and evaluating generalist agents
|
63 |
+
for the web that can follow language instructions to complete complex tasks on any website. In this dataset, we align each HTML document in the dataset with
|
64 |
+
its corresponding webpage screenshot image from the Mind2Web raw dump. This multimodal version addresses the inconvenience of loading images from the ~300GB Mind2Web Raw Dump.
|
65 |
|
66 |
## Dataset Structure
|
67 |
|
|
|
71 |
- test_website: 1019 actions from 142 tasks. Websites are not seen during training.
|
72 |
- test_domain: 4060 actions from 694 tasks. Entire domains are not seen during training.
|
73 |
|
74 |
+
The **_train_** set may include some screenshot images not properly rendered caused by rendering issues during Mind2Web annotation. The three **_test splits (test_task, test_website, test_domain)_** have undergone human verification to confirm element visibility and correct rendering for action prediction.
|
75 |
|
|
|
76 |
|
77 |
### Data Fields
|
78 |
+
Each line in the dataset is an action consisting of screenshot image, HTML text and other fields required for action prediction, for the convenience of inference.
|
79 |
- "annotation_id" (str): unique id for each task
|
80 |
- "website" (str): website name
|
81 |
- "domain" (str): website domain
|
82 |
- "subdomain" (str): website subdomain
|
83 |
- "confirmed_task" (str): task description
|
84 |
- "action_reprs" (list[str]): human readable string representation of the action sequence
|
85 |
+
- **"screenshot" (str): path to the webpage screenshot image corresponding to the HTML.**
|
86 |
- "action_uid" (str): unique id for each action (step)
|
87 |
- "raw_html" (str): raw html of the page before the action is performed
|
88 |
- "cleaned_html" (str): cleaned html of the page before the action is performed
|