JLee0 commited on
Commit
ec1e768
·
verified ·
1 Parent(s): daa1023

Add contributor and clarify manual construction and domain statistics

Browse files
Files changed (1) hide show
  1. README.md +23 -41
README.md CHANGED
@@ -55,17 +55,13 @@ releases in the suite below. The estimate is based on two annotators working
55
  four hours per day, five days per week, for approximately three months
56
  (about twelve weeks).
57
 
58
- ## 🧩 KORMo-VL Benchmark Suite
59
-
60
- This dataset is part of the KORMo-VL Korean GUI-agent benchmark suite. Each
61
- release is built from manually collected Korean data and follows the protocol
62
- of an established international benchmark.
63
-
64
- | Dataset | Setting | Follows |
65
- | --- | --- | --- |
66
- | [Ko-Screenspot](https://huggingface.co/datasets/KORMo-VL/Ko-Screenspot) | Single-image GUI grounding | [ScreenSpot](https://aclanthology.org/2024.acl-long.505/) |
67
- | [Ko-MultimodalMind2Web](https://huggingface.co/datasets/KORMo-VL/Ko-MultimodalMind2Web) | Offline next-action prediction | [Multimodal Mind2Web](https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web) |
68
- | [Ko-Online-Mind2Web](https://huggingface.co/datasets/KORMo-VL/Ko-Online-Mind2Web) | Live task completion | [Online Mind2Web](https://github.com/OSU-NLP-Group/Online-Mind2Web) |
69
 
70
  ## How the benchmark was constructed and verified
71
 
@@ -94,17 +90,7 @@ of an established international benchmark.
94
  1.0 for element accuracy, step success, and task success using the included
95
  scorer.
96
 
97
- ## Relationship to established benchmark formats
98
-
99
- The benchmark follows the offline task/trajectory/action organization used by
100
- [Mind2Web](https://arxiv.org/abs/2306.06070) and
101
- [Multimodal Mind2Web / SeeAct](https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web):
102
- language instructions and screenshots are paired with next-action targets.
103
- The Korean task demonstrations and annotations were collected manually.
104
- The released operation inventory, action history, and accepted alternatives
105
- are explicitly retained so evaluation is reproducible on this collection.
106
-
107
- ## Dataset statistics
108
 
109
  ### Recorded operation distribution
110
 
@@ -125,6 +111,18 @@ are excluded from step-level scoring, leaving 12,905 scored interaction steps.
125
  `statistics.json` includes per-website counts, task categories, trajectory-length
126
  statistics, and alternative-action counts.
127
 
 
 
 
 
 
 
 
 
 
 
 
 
128
  ## Download and preview
129
 
130
  This repository is private. Authenticate with a Hugging Face account that has
@@ -161,25 +159,9 @@ tar -xzf multimodal_mind2web_ko.tar.gz
161
 
162
  ## Evaluation
163
 
164
- - **Element accuracy:** target correctness on 7,580 steps with a primary or
165
- accepted alternative target box. The reference scorer also accepts a valid
166
- location-free alternative such as Enter when that alternative completes the action.
167
- - **Step success rate:** operation, value where required, and target correctness
168
- on all 12,905 scored steps. Each prediction is matched against a complete
169
- accepted answer candidate.
170
- - **Task success rate:** the fraction of 549 tasks for which every scored step
171
- succeeds.
172
-
173
- ```bash
174
- python evaluate.py --data multimodal_mind2web_ko --pred predictions.jsonl
175
- python evaluate.py --data multimodal_mind2web_ko --pred predictions.jsonl --ignore-alt
176
- ```
177
-
178
- Predictions are keyed by `sample_id` and contain `operation`, optional `value`,
179
- and `point` or `bbox` where applicable. The supplied scorer uses exact matching
180
- for action values after text normalization; report any additional semantic
181
- judging policy used for free-text entry when comparing experiments.
182
-
183
 
184
  ## Contact
185
 
 
55
  four hours per day, five days per week, for approximately three months
56
  (about twelve weeks).
57
 
58
+ The benchmark follows the offline task/trajectory/action organization used by
59
+ [Mind2Web](https://arxiv.org/abs/2306.06070) and
60
+ [Multimodal Mind2Web / SeeAct](https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web):
61
+ language instructions and screenshots are paired with next-action targets.
62
+ The Korean task demonstrations and annotations were collected manually.
63
+ The released operation inventory, action history, and accepted alternatives
64
+ are explicitly retained so evaluation is reproducible on this collection.
 
 
 
 
65
 
66
  ## How the benchmark was constructed and verified
67
 
 
90
  1.0 for element accuracy, step success, and task success using the included
91
  scorer.
92
 
93
+ ## 📊 Dataset statistics
 
 
 
 
 
 
 
 
 
 
94
 
95
  ### Recorded operation distribution
96
 
 
111
  `statistics.json` includes per-website counts, task categories, trajectory-length
112
  statistics, and alternative-action counts.
113
 
114
+ ## 🧩 KORMo-VL Benchmark Suite
115
+
116
+ This dataset is part of the KORMo-VL Korean GUI-agent benchmark suite. Each
117
+ release is built from manually collected Korean data and follows the protocol
118
+ of an established international benchmark.
119
+
120
+ | Dataset | Setting | Follows |
121
+ | --- | --- | --- |
122
+ | [Ko-Screenspot](https://huggingface.co/datasets/KORMo-VL/Ko-Screenspot) | Single-image GUI grounding | [ScreenSpot](https://aclanthology.org/2024.acl-long.505/) |
123
+ | [Ko-MultimodalMind2Web](https://huggingface.co/datasets/KORMo-VL/Ko-MultimodalMind2Web) | Offline next-action prediction | [Multimodal Mind2Web](https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web) |
124
+ | [Ko-Online-Mind2Web](https://huggingface.co/datasets/KORMo-VL/Ko-Online-Mind2Web) | Live task completion | [Online Mind2Web](https://github.com/OSU-NLP-Group/Online-Mind2Web) |
125
+
126
  ## Download and preview
127
 
128
  This repository is private. Authenticate with a Hugging Face account that has
 
159
 
160
  ## Evaluation
161
 
162
+ Use the included `evaluate.py` to report element accuracy, step success rate,
163
+ and task success rate on the released scored steps. The reference scorer
164
+ supports accepted alternative actions.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
165
 
166
  ## Contact
167