Datasets:

Modalities:
Image
Text
Formats:
parquet
Languages:
English
ArXiv:
Libraries:
Datasets
Dask
License:
boyuanzheng010 commited on
Commit
305a151
1 Parent(s): c406d5a

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +25 -23
README.md CHANGED
@@ -60,6 +60,16 @@ Multimodal-Mind2Web is the multimodal version of [Mind2Web](https://osu-nlp-grou
60
 
61
  ## Dataset Structure
62
 
 
 
 
 
 
 
 
 
 
 
63
 
64
  ### Data Fields
65
 
@@ -69,30 +79,22 @@ Multimodal-Mind2Web is the multimodal version of [Mind2Web](https://osu-nlp-grou
69
  - "subdomain" (str): website subdomain
70
  - "confirmed_task" (str): task description
71
  - "action_reprs" (list[str]): human readable string representation of the action sequence
72
- - "actions" (list[dict]): list of actions (steps) to complete the task
73
- - - "screenshot" (str): path to the webpage screenshot image corresponding to the HTML.
74
- - "action_uid" (str): unique id for each action (step)
75
- - "raw_html" (str): raw html of the page before the action is performed
76
- - "cleaned_html" (str): cleaned html of the page before the action is performed
77
- - "operation" (dict): operation to perform
78
- - "op" (str): operation type, one of CLICK, TYPE, SELECT
79
- - "original_op" (str): original operation type, contain additional HOVER and ENTER that are mapped to CLICK, not used
80
- - "value" (str): optional value for the operation, e.g., text to type, option to select
81
- - "pos_candidates" (list[dict]): ground truth elements. Here we only include positive elements that exist in "cleaned_html" after our preprocessing, so "pos_candidates" might be empty. The original labeled element can always be found in the "raw_html".
82
- - "tag" (str): tag of the element
83
- - "is_original_target" (bool): whether the element is the original target labeled by the annotator
84
- - "is_top_level_target" (bool): whether the element is a top level target find by our algorithm. please see the paper for more details.
85
- - "backend_node_id" (str): unique id for the element
86
- - "attributes" (str): serialized attributes of the element, use `json.loads` to convert back to dict
87
- - "neg_candidates" (list[dict]): other candidate elements in the page after preprocessing, has similar structure as "pos_candidates"
88
 
89
- ### Data Splits
90
-
91
- - train: 1,009 instances
92
- - test:
93
- - Cross Task: 177 instances, tasks from the same website are seen during training
94
- - Cross Website: 142 instances, websites are not seen during training
95
- - Cross Domain: 694 instances, entire domains are not seen during training
96
 
97
 
98
  ### Disclaimer
 
60
 
61
  ## Dataset Structure
62
 
63
+ ### Data Splits
64
+ - train: 1009 tasks.
65
+ - test:
66
+ - Cross Task: 1339 actions from 177 tasks. Tasks from the same website are seen during training.
67
+ - Cross Website: 1019 actions from 142 tasks. Websites are not seen during training.
68
+ - Cross Domain: 4060 actions from 694 tasks. Entire domains are not seen during training.
69
+
70
+ The three test set splits have went through human verification to confirm element visibility and correct rendering for action prediction.
71
+
72
+ 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.
73
 
74
  ### Data Fields
75
 
 
79
  - "subdomain" (str): website subdomain
80
  - "confirmed_task" (str): task description
81
  - "action_reprs" (list[str]): human readable string representation of the action sequence
82
+ - "screenshot" (str): path to the webpage screenshot image corresponding to the HTML.
83
+ - "action_uid" (str): unique id for each action (step)
84
+ - "raw_html" (str): raw html of the page before the action is performed
85
+ - "cleaned_html" (str): cleaned html of the page before the action is performed
86
+ - "operation" (dict): operation to perform
87
+ - "op" (str): operation type, one of CLICK, TYPE, SELECT
88
+ - "original_op" (str): original operation type, contain additional HOVER and ENTER that are mapped to CLICK, not used
89
+ - "value" (str): optional value for the operation, e.g., text to type, option to select
90
+ - "pos_candidates" (list[dict]): ground truth elements. Here we only include positive elements that exist in "cleaned_html" after our preprocessing, so "pos_candidates" might be empty. The original labeled element can always be found in the "raw_html".
91
+ - "tag" (str): tag of the element
92
+ - "is_original_target" (bool): whether the element is the original target labeled by the annotator
93
+ - "is_top_level_target" (bool): whether the element is a top level target find by our algorithm. please see the paper for more details.
94
+ - "backend_node_id" (str): unique id for the element
95
+ - "attributes" (str): serialized attributes of the element, use `json.loads` to convert back to dict
96
+ - "neg_candidates" (list[dict]): other candidate elements in the page after preprocessing, has similar structure as "pos_candidates"
 
97
 
 
 
 
 
 
 
 
98
 
99
 
100
  ### Disclaimer