JLee0 commited on
Commit
a08d02b
·
verified ·
1 Parent(s): c6bf047

Release manually annotated Korean benchmark with construction details and statistics

Browse files
README.md ADDED
@@ -0,0 +1,263 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pretty_name: Ko-MultimodalMind2Web
3
+ license: cc-by-4.0
4
+ language:
5
+ - ko
6
+ task_categories:
7
+ - image-text-to-text
8
+ size_categories:
9
+ - 10K<n<100K
10
+ tags:
11
+ - korean
12
+ - gui-agent
13
+ - web-agent
14
+ - human-annotated
15
+ - evaluation
16
+ modalities:
17
+ - image
18
+ - text
19
+ configs:
20
+ - config_name: default
21
+ data_files:
22
+ - split: test
23
+ path: examples/test.parquet
24
+ ---
25
+
26
+ # Ko-MultimodalMind2Web
27
+
28
+ **A Korean offline web-agent benchmark built from browser demonstrations and
29
+ annotations created manually by Korean human annotators.** The release contains
30
+ **549 tasks, 12,905 scored action steps, and 144 websites**.
31
+ The shared Korean annotation project involved two annotators working four
32
+ hours per day, five days per week, for approximately three months: **about
33
+ 480 annotator-hours of combined collection and annotation effort**.
34
+
35
+ The model receives a Korean task, the current screenshot, and preceding action
36
+ history, then predicts the next action and its target or value.
37
+
38
+ ## Human annotation and collection effort
39
+
40
+ **The Korean data was built manually by two Korean human annotators.** They
41
+ designed realistic tasks, operated the websites themselves, and produced browser
42
+ demonstrations and annotations from those interactions. This human collection
43
+ is the basis of the Korean benchmark; the published task interfaces follow the
44
+ established international benchmark protocols described below.
45
+
46
+ | Collection item | Reported schedule |
47
+ | --- | --- |
48
+ | Korean human annotators | 2 |
49
+ | Work per annotator per day | 4 hours |
50
+ | Working days per week | 5 |
51
+ | Annotation period | Approximately 3 months |
52
+ | Estimated effort per annotator | Approximately 240 hours |
53
+ | Estimated combined effort | **Approximately 480 annotator-hours** |
54
+
55
+ The effort estimate uses **2 × 4 hours × 5 days × approximately 12 weeks**.
56
+ Three months is approximated as 12 working weeks for this estimate. This is the
57
+ shared collection and annotation effort underpinning the three Korean releases:
58
+ grounding, offline action prediction, and online task completion.
59
+
60
+ ## How the benchmark was constructed and verified
61
+
62
+ 1. **Human task design and demonstration.** Korean annotators created realistic
63
+ tasks on Korean websites and Korean-language interfaces, then performed the
64
+ complete tasks themselves. Browser events, screenshots, and target-element
65
+ information were captured from their actual interactions.
66
+ 2. **Step-level annotation.** The recorded demonstrations were organized into
67
+ task-linked steps with an operation, argument/value, target geometry when
68
+ relevant, and the preceding action history. Both original and concise
69
+ task-instruction variants are supplied.
70
+ 3. **Manual target correction.** Native browser dialogs, separate popup windows,
71
+ operating-system select menus, and dynamically replaced controls were
72
+ inspected using their screenshots. Missing or incorrectly aligned targets
73
+ were corrected manually. Selection actions were checked against subsequent
74
+ screenshots to distinguish the selected value from a transient highlight.
75
+ 4. **Alternative-answer review.** Equivalent form-submission actions were
76
+ represented as alternative valid answers. The final release includes
77
+ **350 manually located alternative CLICK targets**, together with
78
+ template-matched alternatives and valid Enter-key submissions. Operation,
79
+ value, and location must belong to the same accepted answer candidate.
80
+ 5. **Consistency and scoring checks.** All 12,905 scored-step IDs are unique;
81
+ all steps resolve to one of the 549 tasks and agree with task-level step
82
+ counts. All 7,197 primary target boxes are finite, have positive area, and
83
+ lie within their screenshot dimensions. Gold-label round-trip checks return
84
+ 1.0 for element accuracy, step success, and task success using the included
85
+ scorer.
86
+
87
+ ## Relationship to established benchmark formats
88
+
89
+ The benchmark follows the offline task/trajectory/action organization used by
90
+ [Mind2Web](https://arxiv.org/abs/2306.06070) and
91
+ [Multimodal Mind2Web / SeeAct](https://huggingface.co/datasets/osunlp/Multimodal-Mind2Web):
92
+ language instructions and screenshots are paired with next-action targets.
93
+ The Korean task demonstrations and annotations were collected manually.
94
+ The released operation inventory, action history, and accepted alternatives
95
+ are explicitly retained so evaluation is reproducible on this collection.
96
+
97
+ ## Dataset statistics
98
+
99
+ | Statistic | Full release |
100
+ | --- | --- |
101
+ | Human-recorded tasks | 549 |
102
+ | Websites | 144 |
103
+ | Recorded steps in reference-task metadata | 14,042 |
104
+ | Distributed scored steps / screenshots | 12,905 |
105
+ | Mean / median scored steps per task | 23.51 / 21 |
106
+ | Scored steps per task, minimum–maximum | 3–100 |
107
+ | Steps with primary target boxes | 7,197 |
108
+ | Steps with a primary or alternative target box | 7,580 |
109
+ | Steps with alternative valid actions | 755 |
110
+ | Alternative valid actions | 756 |
111
+ | Original / concise instruction records | 549 / 549 |
112
+ | Viewer examples | 10 |
113
+
114
+ ### Recorded operation distribution
115
+
116
+ | Category | Count | Share |
117
+ | --- | --- | --- |
118
+ | CLICK | 5,879 | 45.56% |
119
+ | TYPE | 1,399 | 10.84% |
120
+ | HOTKEY | 444 | 3.44% |
121
+ | SCROLL | 5,007 | 38.80% |
122
+ | CLEAR | 15 | 0.12% |
123
+ | GO_BACK | 134 | 1.04% |
124
+ | DRAG | 27 | 0.21% |
125
+
126
+ The complete human reference recordings contain 14,042 steps. The distributed
127
+ step file contains the 12,905 scored steps; `GOTO`, `WAIT`, and `FINISH` are
128
+ excluded from step scoring. `statistics.json` includes per-website counts,
129
+ task categories, trajectory-length statistics, and alternative-action counts.
130
+
131
+ ## Files and fields
132
+
133
+ | File | Contents |
134
+ | --- | --- |
135
+ | `multimodal_mind2web_ko.tar.gz` | Scored steps, task metadata, instruction variants, and screenshots |
136
+ | `examples/test.parquet` | Ten viewer examples with embedded images |
137
+ | `examples/samples.jsonl` | Complete example records |
138
+ | `examples/images/` | The corresponding example PNGs |
139
+ | `statistics.json` | Full-release aggregate statistics |
140
+ | `evaluate.py` | Reference element/step/task scorer |
141
+
142
+ Extraction creates `multimodal_mind2web_ko/steps.jsonl`, `tasks.jsonl`,
143
+ `tasks_rewrite.jsonl`, and `images/`. The step records include `sample_id`,
144
+ `task_id`, `instruction`, `image`, `image_size`, `action`, `previous_actions`,
145
+ `scored`, target geometry, and `alt_actions`. Image paths are relative to that
146
+ extracted directory. The preview exposes the structured action as `action_json`
147
+ and places `task_id` in the final column.
148
+
149
+ Pixel coordinates, [0, 1] normalized fields, and the original `*_999` grid
150
+ fields are preserved. The scorer evaluates in the 0–999 grid, converting
151
+ pixel or [0, 1] predictions to that coordinate system. Use the same convention
152
+ when comparing scores.
153
+
154
+ ## Download and preview
155
+
156
+ This repository is private. Authenticate with a Hugging Face account that has
157
+ access to the repository.
158
+
159
+ ```python
160
+ from huggingface_hub import snapshot_download
161
+
162
+ snapshot_download(
163
+ "KORMo-VL/Ko-MultimodalMind2Web",
164
+ repo_type="dataset",
165
+ local_dir="Ko-MultimodalMind2Web",
166
+ allow_patterns=["README.md", "statistics.json", "evaluate.py", "multimodal_mind2web_ko.tar.gz"],
167
+ token=True,
168
+ )
169
+ ```
170
+
171
+ The dataset viewer contains **10 illustrative examples**. The complete benchmark
172
+ is provided in the separate full-data file listed above. These examples are
173
+ selected from the full data.
174
+
175
+ ```python
176
+ from datasets import load_dataset
177
+
178
+ preview = load_dataset("KORMo-VL/Ko-MultimodalMind2Web", split="test", token=True)
179
+ ```
180
+
181
+ Extract the full data from the downloaded repository directory:
182
+
183
+ ```bash
184
+ cd Ko-MultimodalMind2Web
185
+ tar -xzf multimodal_mind2web_ko.tar.gz
186
+ ```
187
+
188
+ ## Evaluation
189
+
190
+ - **Element accuracy:** target correctness on 7,580 steps with a primary or
191
+ accepted alternative target box. The reference scorer also accepts a valid
192
+ location-free alternative such as Enter when that alternative completes the action.
193
+ - **Step success rate:** operation, value where required, and target correctness
194
+ on all 12,905 scored steps. Each prediction is matched against a complete
195
+ accepted answer candidate.
196
+ - **Task success rate:** the fraction of 549 tasks for which every scored step
197
+ succeeds.
198
+
199
+ ```bash
200
+ python evaluate.py --data multimodal_mind2web_ko --pred predictions.jsonl
201
+ python evaluate.py --data multimodal_mind2web_ko --pred predictions.jsonl --ignore-alt
202
+ ```
203
+
204
+ Predictions are keyed by `sample_id` and contain `operation`, optional `value`,
205
+ and `point` or `bbox` where applicable. The supplied scorer uses exact matching
206
+ for action values after text normalization; report any additional semantic
207
+ judging policy used for free-text entry when comparing experiments.
208
+
209
+ ## Related Korean benchmark releases
210
+
211
+ - [Ko-Screenspot](https://huggingface.co/datasets/KORMo-VL/Ko-Screenspot): single-image GUI grounding.
212
+ - [Ko-Online-Mind2Web](https://huggingface.co/datasets/KORMo-VL/Ko-Online-Mind2Web): live website task completion.
213
+
214
+ ## Contact
215
+
216
+ * KyungTae Lim, Professor at KAIST. `ktlim@kaist.ac.kr`
217
+
218
+ ## Contributor
219
+
220
+ * Yerin Nam ([yerinnam@kaist.ac.kr](mailto:yerinnam@kaist.ac.kr))
221
+ * Minkyung Cho ([minkyung.cho@kaist.ac.kr](mailto:minkyung.cho@kaist.ac.kr))
222
+ * Minkyu Kim ([minkyukim@kaist.ac.kr](mailto:minkyukim@kaist.ac.kr))
223
+ * Seungwoo Song ([sswoo@kaist.ac.kr](mailto:sswoo@kaist.ac.kr))
224
+ * Hangyeol Yoo ([hgyoo@kaist.ac.kr](mailto:hgyoo@kaist.ac.kr))
225
+ * KyungTae Lim ([ktlim@kaist.ac.kr](mailto:ktlim@kaist.ac.kr))
226
+
227
+ ## Citation
228
+
229
+ References for KORMo and the benchmark protocol followed by this release:
230
+
231
+ ```bibtex
232
+ @misc{KORMo,
233
+ author = {Minjun Kim and Hyeonseok Lim and Hangyeol Yoo and Inho Won and Seungwoo Song and Minkyung Cho and Junghun Yuk and Changsu Choi and Dongjae Shin and Huije Lee and Hoyun Song and Alice Oh and KyungTae Lim},
234
+ title = {KORMo: Korean Open Reasoning Model for Everyone},
235
+ year = {2025},
236
+ publisher = {GitHub},
237
+ journal = {Technical Report},
238
+ url = {https://arxiv.org/abs/2510.09426}
239
+ }
240
+ ```
241
+
242
+ ```bibtex
243
+ @inproceedings{zheng2024seeact,
244
+ title = {GPT-4V(ision) is a Generalist Web Agent, if Grounded},
245
+ author = {Boyuan Zheng and Boyu Gou and Jihyung Kil and Huan Sun and Yu Su},
246
+ booktitle = {Forty-first International Conference on Machine Learning},
247
+ year = {2024},
248
+ url = {https://openreview.net/forum?id=piecKJ2DlB}
249
+ }
250
+ ```
251
+
252
+ ```bibtex
253
+ @inproceedings{deng2023mind2web,
254
+ title = {Mind2Web: Towards a Generalist Agent for the Web},
255
+ author = {Xiang Deng and Yu Gu and Boyuan Zheng and Shijie Chen and Sam Stevens and Boshi Wang and Huan Sun and Yu Su},
256
+ booktitle = {Advances in Neural Information Processing Systems},
257
+ volume = {36},
258
+ pages = {28091--28114},
259
+ year = {2023},
260
+ publisher = {Curran Associates, Inc.},
261
+ url = {https://proceedings.neurips.cc/paper_files/paper/2023/file/5950bf290a1570ea401bf98882128160-Paper-Datasets_and_Benchmarks.pdf}
262
+ }
263
+ ```
evaluate.py ADDED
@@ -0,0 +1,333 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Score model predictions against the Korean Multimodal Mind2Web sets.
2
+
3
+ Predictions are a JSONL file keyed by ``sample_id``::
4
+
5
+ {"sample_id": "0714_Kaggle_1#2", "operation": "CLICK", "point": [0.78, 0.03]}
6
+ {"sample_id": "0714_Kaggle_1#5", "operation": "TYPE", "bbox": [0.1, 0.2, 0.3, 0.25],
7
+ "value": "covid-19"}
8
+
9
+ Coordinates may be given either normalised to [0, 1] or in screenshot pixels;
10
+ the scorer detects which by checking whether every value is <= 1.
11
+
12
+ Metrics
13
+ -------
14
+ All three are scored over the same ``steps/samples.jsonl``; they differ only in
15
+ what they demand of a step.
16
+
17
+ element accuracy the target element was hit — predicted point inside the gold
18
+ box (default) or IoU >= --iou. Only steps with
19
+ ``has_element`` count towards it.
20
+ step success the element was hit AND the operation and its value are right.
21
+ Only steps with ``scored`` count — WAIT and FINISH are left
22
+ out, since neither is an agent decision about the page.
23
+ task success every scored step of the task counts as a step success
24
+
25
+ More than one right answer
26
+ --------------------------
27
+ Some steps have two. A typed form submits either by clicking its button or by
28
+ pressing Enter, and the annotator recorded whichever they happened to do; the
29
+ other is in ``alt_actions`` and is just as correct. A prediction counts if it
30
+ matches **any** of them — the recorded action or an alternative — and each
31
+ alternative brings its own box, so a click on the submit button is scored
32
+ against that button, not against whatever the recorded action pointed at.
33
+
34
+ Pass ``--ignore-alt`` to score only the recorded action (the pre-alt_actions
35
+ numbers, for comparison).
36
+
37
+ Usage:
38
+ python evaluate.py --pred preds.jsonl
39
+ python evaluate.py --pred preds.jsonl --match iou --iou 0.5 --by-website
40
+ python evaluate.py --pred preds.jsonl --ignore-alt
41
+ python evaluate.py --pred preds.jsonl --type-verdicts type_verdicts.jsonl
42
+ """
43
+
44
+ from __future__ import annotations
45
+
46
+ import argparse
47
+ import json
48
+ import os
49
+ import re
50
+ from collections import defaultdict
51
+ from typing import Any
52
+
53
+ # Operations whose textual argument is compared when scoring step success.
54
+ VALUE_OPS = {"TYPE", "SCROLL", "HOTKEY", "GOTO"}
55
+
56
+
57
+ def load_jsonl(path: str) -> list[dict]:
58
+ rows = []
59
+ with open(path, encoding="utf-8") as fh:
60
+ for line in fh:
61
+ line = line.strip()
62
+ if line:
63
+ rows.append(json.loads(line))
64
+ return rows
65
+
66
+
67
+ def norm_text(value: Any) -> str:
68
+ if value is None:
69
+ return ""
70
+ return re.sub(r"\s+", " ", str(value)).strip().lower()
71
+
72
+
73
+ GRID = 999
74
+
75
+
76
+ def to_grid999(coords, size):
77
+ """예측 좌표 -> 0-999 그리드. ★채점은 전부 이 좌표계에서 한다★.
78
+
79
+ 골든이 `gold_bbox_999`(0-999 정수)이므로 예측도 같은 격자로 옮겨 비교한다.
80
+ 전부 <= 1.0 -> 0-1 정규화 -> x 999
81
+ 그 외 -> 스크린샷 픽셀 -> / image_size x 999
82
+ """
83
+ if coords is None:
84
+ return None
85
+ coords = [float(c) for c in coords]
86
+ if all(c <= 1.0 for c in coords):
87
+ return [c * GRID for c in coords]
88
+ w, h = size
89
+ return [c / (w if i % 2 == 0 else h) * GRID for i, c in enumerate(coords)]
90
+
91
+
92
+ def point_in_box(point, box) -> bool:
93
+ x, y = point
94
+ return box[0] <= x <= box[2] and box[1] <= y <= box[3]
95
+
96
+
97
+ def iou(a, b) -> float:
98
+ ix1, iy1 = max(a[0], b[0]), max(a[1], b[1])
99
+ ix2, iy2 = min(a[2], b[2]), min(a[3], b[3])
100
+ if ix2 <= ix1 or iy2 <= iy1:
101
+ return 0.0
102
+ inter = (ix2 - ix1) * (iy2 - iy1)
103
+ area_a = (a[2] - a[0]) * (a[3] - a[1])
104
+ area_b = (b[2] - b[0]) * (b[3] - b[1])
105
+ return inter / (area_a + area_b - inter)
106
+
107
+
108
+ def candidates(gold: dict, use_alt: bool = True) -> list[dict]:
109
+ """Every action that counts as correct for this step, recorded one first.
110
+
111
+ Each entry is flattened to ``{operation, value, gold_bbox}`` so the scoring
112
+ below does not have to care whether it came from ``action`` or from
113
+ ``alt_actions``.
114
+ """
115
+ out = [{"operation": gold["action"]["operation"],
116
+ "value": gold["action"].get("value"),
117
+ # ★0-999 그리드 골든★ — 채점 좌표계가 픽셀에서 그리드로 바뀌었다.
118
+ "gold_bbox": gold.get("gold_bbox_999")}]
119
+ if use_alt:
120
+ for alt in gold.get("alt_actions") or []:
121
+ if alt.get("status") not in (None, "ready"):
122
+ continue # 좌표가 아직 없는 대안은 채점에 넣지 않는다
123
+ out.append({"operation": alt["operation"], "value": alt.get("value"),
124
+ "gold_bbox": alt.get("gold_bbox_999")})
125
+ return out
126
+
127
+
128
+ def hits_box(pred: dict, box, size, match: str, iou_thr: float) -> bool:
129
+ """Does the prediction land on this box?"""
130
+ if box is None:
131
+ return False
132
+ if match == "iou":
133
+ pbox = to_grid999(pred.get("bbox"), size)
134
+ return bool(pbox) and iou(pbox, box) >= iou_thr
135
+
136
+ # default: "point" — a predicted point (or the centre of a predicted box)
137
+ # must fall inside the box. 둘 다 0-999 그리드다.
138
+ point = to_grid999(pred.get("point"), size)
139
+ if point is None:
140
+ pbox = to_grid999(pred.get("bbox"), size)
141
+ if pbox is None:
142
+ return False
143
+ point = [(pbox[0] + pbox[2]) / 2, (pbox[1] + pbox[3]) / 2]
144
+ return point_in_box(point, box)
145
+
146
+
147
+ def matches_actionless(pred: dict, gold: dict, use_alt: bool = True,
148
+ type_ok: bool | None = None) -> bool:
149
+ """Did the prediction pick a correct answer that has no place on screen?
150
+
151
+ Pressing Enter to submit a form is one of the two right answers on 388
152
+ steps, and it points at nothing. It is still a right answer, so it counts
153
+ as a hit — the step asked "how do you submit this form", and both answers
154
+ do it.
155
+ """
156
+ for c in candidates(gold, use_alt):
157
+ if c["gold_bbox"] is not None:
158
+ continue
159
+ if norm_text(pred.get("operation")) != norm_text(c["operation"]):
160
+ continue
161
+ if c["operation"] in VALUE_OPS \
162
+ and norm_text(pred.get("value")) != norm_text(c["value"]):
163
+ # 완전 일치는 실패했지만 LLM judge 가 같은 뜻이라고 본 TYPE 이면 통과시킨다
164
+ # (eval_type.py 참고). judge 는 구제만 하고 통과한 것을 떨어뜨리지 않는다.
165
+ if not (type_ok and c["operation"] == "TYPE"):
166
+ continue
167
+ return True
168
+ return False
169
+
170
+
171
+ def element_hit(pred: dict, gold: dict, match: str, iou_thr: float,
172
+ use_alt: bool = True, type_ok: bool | None = None) -> bool:
173
+ """True when the prediction picks any acceptable answer for this step.
174
+
175
+ Usually that means landing on one of the accepted boxes. Where an accepted
176
+ answer has no box at all — Enter submits the form just as the button does —
177
+ naming that action counts too; there is nothing to land on.
178
+ """
179
+ size = gold["image_size"]
180
+ if any(hits_box(pred, c["gold_bbox"], size, match, iou_thr)
181
+ for c in candidates(gold, use_alt)):
182
+ return True
183
+ return matches_actionless(pred, gold, use_alt, type_ok)
184
+
185
+
186
+ def grounded(gold: dict, use_alt: bool = True) -> bool:
187
+ """Is there any correct answer for this step that has a box?
188
+
189
+ ``has_element`` answers this for the recorded action only. A step whose
190
+ gold is Enter but whose alternative is "click the submit button" does have
191
+ a right answer with coordinates, so a model that clicks it should get the
192
+ credit.
193
+ """
194
+ return any(c["gold_bbox"] is not None for c in candidates(gold, use_alt))
195
+
196
+
197
+ def step_hit(pred: dict, gold: dict, match: str, iou_thr: float,
198
+ use_alt: bool = True, type_ok: bool | None = None) -> bool:
199
+ """True when the prediction matches one of the step's correct actions.
200
+
201
+ Operation, value and element are checked against the *same* candidate — a
202
+ prediction cannot borrow the operation from one answer and the box from
203
+ another.
204
+ """
205
+ size = gold["image_size"]
206
+ for c in candidates(gold, use_alt):
207
+ if norm_text(pred.get("operation")) != norm_text(c["operation"]):
208
+ continue
209
+ if c["operation"] in VALUE_OPS \
210
+ and norm_text(pred.get("value")) != norm_text(c["value"]):
211
+ # 완전 일치는 실패했지만 LLM judge 가 같은 뜻이라고 본 TYPE 이면 통과시킨다
212
+ # (eval_type.py 참고). judge 는 구제만 하고 통과한 것을 떨어뜨리지 않는다.
213
+ if not (type_ok and c["operation"] == "TYPE"):
214
+ continue
215
+ if c["gold_bbox"] is not None \
216
+ and not hits_box(pred, c["gold_bbox"], size, match, iou_thr):
217
+ continue
218
+ return True
219
+ return False
220
+
221
+
222
+ def summarise(counts: dict[str, list[int]]) -> dict:
223
+ return {
224
+ key: {
225
+ "correct": hit,
226
+ "total": total,
227
+ "score": round(hit / total, 4) if total else None,
228
+ }
229
+ for key, (hit, total) in sorted(counts.items())
230
+ }
231
+
232
+
233
+ def main() -> None:
234
+ here = os.path.dirname(os.path.abspath(__file__))
235
+ parser = argparse.ArgumentParser()
236
+ parser.add_argument("--pred", required=True, help="JSONL of predictions")
237
+ parser.add_argument("--data", default=here, help="dataset root")
238
+ parser.add_argument("--match", choices=["point", "iou"], default="point")
239
+ parser.add_argument("--iou", type=float, default=0.5)
240
+ parser.add_argument("--by-website", action="store_true")
241
+ parser.add_argument("--ignore-alt", action="store_true",
242
+ help="score only the recorded action, ignoring alt_actions")
243
+ parser.add_argument("--type-verdicts", default=None,
244
+ help="eval_type.py 가 낸 TYPE 판정 JSONL — 완전 일치 대신 쓴다")
245
+ parser.add_argument("--out", default=None, help="write the report to this JSON file")
246
+ opts = parser.parse_args()
247
+
248
+ preds = {p["sample_id"]: p for p in load_jsonl(opts.pred)}
249
+ type_ok = {}
250
+ if opts.type_verdicts:
251
+ type_ok = {v["sample_id"]: bool(v.get("correct"))
252
+ for v in load_jsonl(opts.type_verdicts)}
253
+ step_gold = load_jsonl(os.path.join(opts.data, "steps.jsonl"))
254
+ task_gold = load_jsonl(os.path.join(opts.data, "tasks.jsonl"))
255
+
256
+ el_hit = el_total = el_na = 0
257
+ el_by_site: dict[str, list[int]] = defaultdict(lambda: [0, 0])
258
+ step_ok: dict[str, bool] = {}
259
+ step_by_op: dict[str, list[int]] = defaultdict(lambda: [0, 0])
260
+
261
+ for gold in step_gold:
262
+ pred = preds.get(gold["sample_id"])
263
+
264
+ # Element accuracy: did the prediction land on the target element?
265
+ # Scored wherever some correct answer has a box — which includes the
266
+ # steps whose *alternative* is the click, not just ``has_element``.
267
+ use_alt = not opts.ignore_alt
268
+ if grounded(gold, use_alt):
269
+ tok = type_ok.get(gold["sample_id"])
270
+ hit = bool(pred) and element_hit(pred, gold, opts.match, opts.iou, use_alt, tok)
271
+ # 좌표 없이 인정된 답(Enter)이 몇 개인지는 따로 세어 리포트에 남긴다
272
+ el_na += bool(hit and pred and matches_actionless(pred, gold, use_alt, tok))
273
+ el_hit += hit
274
+ el_total += 1
275
+ site = el_by_site[gold["website"]]
276
+ site[0] += hit
277
+ site[1] += 1
278
+
279
+ # Step success: the same steps, but the operation and value must match
280
+ # too. WAIT and FINISH are not scored.
281
+ if not gold["scored"]:
282
+ continue
283
+ ok = bool(pred) and step_hit(pred, gold, opts.match, opts.iou,
284
+ not opts.ignore_alt,
285
+ type_ok.get(gold["sample_id"]))
286
+ step_ok[gold["sample_id"]] = ok
287
+ op = step_by_op[gold["action"]["operation"]]
288
+ op[0] += ok
289
+ op[1] += 1
290
+
291
+ task_hit = 0
292
+ for task in task_gold:
293
+ ids = task["scored_step_ids"]
294
+ task_hit += bool(ids) and all(step_ok.get(i, False) for i in ids)
295
+
296
+ n_steps = len(step_ok)
297
+ report = {
298
+ "predictions": len(preds),
299
+ "match": opts.match,
300
+ "iou_threshold": opts.iou if opts.match == "iou" else None,
301
+ "alt_actions": "ignored" if opts.ignore_alt else "accepted",
302
+ "type_value_match": ("llm-judge + exact" if opts.type_verdicts else "exact"),
303
+ "element_accuracy": {
304
+ "correct": el_hit,
305
+ "total": el_total,
306
+ "score": round(el_hit / el_total, 4) if el_total else None,
307
+ # 좌표 없이(Enter 로) 맞은 것으로 인정된 수 — 분모에는 그대로 들어 있다
308
+ "credited_without_coordinates": el_na,
309
+ },
310
+ "step_success_rate": {
311
+ "correct": sum(step_ok.values()),
312
+ "total": n_steps,
313
+ "score": round(sum(step_ok.values()) / n_steps, 4) if n_steps else None,
314
+ },
315
+ "task_success_rate": {
316
+ "correct": task_hit,
317
+ "total": len(task_gold),
318
+ "score": round(task_hit / len(task_gold), 4) if task_gold else None,
319
+ },
320
+ "step_success_by_operation": summarise(step_by_op),
321
+ }
322
+ if opts.by_website:
323
+ report["element_accuracy_by_website"] = summarise(el_by_site)
324
+
325
+ text = json.dumps(report, ensure_ascii=False, indent=2)
326
+ print(text)
327
+ if opts.out:
328
+ with open(opts.out, "w", encoding="utf-8") as fh:
329
+ fh.write(text)
330
+
331
+
332
+ if __name__ == "__main__":
333
+ main()
examples/images/0714_X_2_1/frame_6.png ADDED

Git LFS Details

  • SHA256: 8a44f684d09183ad338a5bc1431666a52973ecc3a75700aecfa810fdb5d1c396
  • Pointer size: 131 Bytes
  • Size of remote file: 223 kB
examples/images/0714_ebsi_1/frame_3.png ADDED

Git LFS Details

  • SHA256: fd93b2b909bd2c59a39566e7a09d1d0a26d123b65ed9130db66aa59db58adf5a
  • Pointer size: 132 Bytes
  • Size of remote file: 1.15 MB
examples/images/0714_google_docs_1/frame_2.png ADDED

Git LFS Details

  • SHA256: b42da540fe11dbe332cf9d576d26bb0dee040b52bb85b9df411fea118f590d51
  • Pointer size: 131 Bytes
  • Size of remote file: 193 kB
examples/images/0714_instagram_2_1/frame_3.png ADDED

Git LFS Details

  • SHA256: 1f36609eed7aa6d544e03d1812e47a496f8fff6c4e54b646bb6672ee903f6f31
  • Pointer size: 131 Bytes
  • Size of remote file: 161 kB
examples/images/0714_nate_1/frame_2.png ADDED

Git LFS Details

  • SHA256: a393d212bea2da3cbe9f050d07b45e7fb7d8c7a7976b7d2ab5947c9184159b49
  • Pointer size: 132 Bytes
  • Size of remote file: 1.1 MB
examples/images/0714_naver_webtoon_1/frame_2.png ADDED

Git LFS Details

  • SHA256: e536d21df9451fad7a0a3bf9ed40ed71f619d47b23a028a0df51d46a2dedc4b3
  • Pointer size: 132 Bytes
  • Size of remote file: 1.09 MB
examples/images/0714_ohou_se_1/frame_3.png ADDED

Git LFS Details

  • SHA256: 6ae72f71343ee1b9c7244e6612610bbc635e7e7ceba89eaebd8f0ac529872b0b
  • Pointer size: 131 Bytes
  • Size of remote file: 156 kB
examples/images/0714_pexels_1/frame_2.png ADDED

Git LFS Details

  • SHA256: 57dc1a1891eac99deeccac1d98e2471b8cc1b52e8cdc6618416d1143d93755c6
  • Pointer size: 131 Bytes
  • Size of remote file: 609 kB
examples/images/0714_pinterest_1/frame_2.png ADDED

Git LFS Details

  • SHA256: 966ff6c958bfa466a2680df23d183f52845984472ba26ffd108913a1d70c0918
  • Pointer size: 131 Bytes
  • Size of remote file: 549 kB
examples/images/0720_zigzag_1/frame_22.png ADDED

Git LFS Details

  • SHA256: 5f639f415ca93e765e8da9bb84e3628a305a22967f2ddcefc989dafbf6ff2006
  • Pointer size: 131 Bytes
  • Size of remote file: 844 kB
examples/samples.jsonl ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {"sample_id": "0714_X_2_1#2", "date": "0714", "website": "X_2", "task_name": "X_2_1", "instruction": "러닝을 시작하려고 하는데 다른 사람들이 어떤 러닝화를 신는지 듣고 싶어. X에서 러닝하는 분들에게 어떤 러닝화를 신는지 묻는 글을 올려줘.", "step_idx": 1, "raw_step": 2, "image": "images/0714_X_2_1/frame_6.png", "image_size": [1280, 826], "action": {"operation": "CLICK", "value": null, "target_text": "이메일 또는 사용자 이름"}, "element": {"tag": "DIV", "text": "이메일 또는 사용자 이름", "aria_label": null, "role": null, "id": null, "class": "jf-element jf-float-label-container", "href": null}, "gold_bbox": [51.5, 619.5, 601.4, 706.8], "gold_bbox_norm": [0.04023, 0.75, 0.46985, 0.8557], "gold_point": [326.45, 663.15], "gold_point_norm": [0.25504, 0.80285], "click_point": [309.26, 658.99], "click_point_norm": [0.24161, 0.7978], "has_element": true, "scored": true, "url": "https://x.com/", "page_title": "X. 무슨 일이 일어나고 있나요? / X", "previous_actions": ["GOTO https://x.com/"], "task_type": "write", "login_required": true, "total_steps": 11, "alt_actions": [], "gold_bbox_999": [40, 749, 469, 855], "gold_point_999": [255, 802], "click_point_999": [241, 797], "mc_candidates": {"A": "CLICK — click(x, y) — 화면의 한 점을 클릭한다", "B": "TYPE — type(x, y, \"text\") — 입력칸에 텍스트를 입력한다", "C": "SCROLL — scroll(\"up\"|\"down\"|\"left\"|\"right\") — 페이지를 한 화면 넘긴다", "D": "HOTKEY — press(\"key\", ...) — 키나 조합키를 누른다", "E": "CLEAR — clear(x, y) — 입력칸의 기존 내용을 지운다", "F": "GO_BACK — go_back() — 이전 페이지로 돌아간다", "G": "DRAG — drag(x1, y1, x2, y2) — 한 점에서 다른 점으로 끈다"}, "mc_answers": ["A"], "mc_answer": "A", "mc_n_options": 7, "mc_needs_point": true, "mc_needs_value": false, "task_id": "0714_X_2_1"}
2
+ {"sample_id": "0720_zigzag_1#20", "date": "0720", "website": "zigzag", "task_name": "zigzag_1", "instruction": "여름에 편하게 입을 검은색 반팔 원피스를 찾아주세요. 지그재그에서 반팔 원피스를 검색하고 색상은 블랙, 가격은 5만 원 이하로 설정해주세요. 구매 후기가 충분하고 무료배송이 가능한 상품 하나를 골라 소재와 총기장, 세탁 방법을 확인한 뒤 찜해주세요", "step_idx": 19, "raw_step": 20, "image": "images/0720_zigzag_1/frame_22.png", "image_size": [1280, 802], "action": {"operation": "CLICK", "value": null, "target_text": null}, "element": {"tag": "A", "text": null, "aria_label": null, "role": null, "id": null, "class": "css-1pjr9xx product-card-link", "href": "https://zigzag.kr/catalog/products/127783711"}, "gold_bbox": [297.8, 191.4, 519.0, 600.7], "gold_bbox_norm": [0.23265, 0.23865, 0.40547, 0.74901], "gold_point": [408.4, 396.05], "gold_point_norm": [0.31906, 0.49383], "click_point": [437.2, 299.69], "click_point_norm": [0.34157, 0.37368], "has_element": true, "scored": true, "url": "https://zigzag.kr/search?keyword=%EB%B0%98%ED%8C%94%20%EC%9B%90%ED%94%BC%EC%8A%A4&only_free_shipping=true", "page_title": "지그재그 스토어", "previous_actions": ["GOTO https://zigzag.kr/", "CLICK", "CLICK 카카오로 3초 만에 시작하기", "CLICK", "TYPE \"[ID]\" into input", "CLICK 비밀번호 입력", "CLICK", "TYPE \"[PASSWORD]\" into input", "HOTKEY ['Enter']", "CLICK 홈", "CLICK", "CLICK", "TYPE \"반팔 원피스\" into input", "HOTKEY ['Enter']", "SCROLL down", "SCROLL up", "CLICK", "SCROLL down", "SCROLL down"], "task_type": "search", "login_required": true, "total_steps": 77, "alt_actions": [], "gold_bbox_999": [232, 238, 405, 748], "gold_point_999": [319, 493], "click_point_999": [341, 373], "mc_candidates": {"A": "CLICK — click(x, y) — 화면의 한 점을 클릭한다", "B": "TYPE — type(x, y, \"text\") — 입력칸에 텍스트를 입력한다", "C": "SCROLL — scroll(\"up\"|\"down\"|\"left\"|\"right\") — 페이지를 한 화면 넘긴다", "D": "HOTKEY — press(\"key\", ...) — 키나 조합키를 누른다", "E": "CLEAR — clear(x, y) — 입력칸의 기존 내용을 지운다", "F": "GO_BACK — go_back() — 이전 페이지로 돌아간다", "G": "DRAG — drag(x1, y1, x2, y2) — 한 점에서 다른 점으로 끈다"}, "mc_answers": ["A"], "mc_answer": "A", "mc_n_options": 7, "mc_needs_point": true, "mc_needs_value": false, "task_id": "0720_zigzag_1"}
3
+ {"sample_id": "0714_ebsi_1#2", "date": "0714", "website": "ebsi", "task_name": "ebsi_1", "instruction": "나에게 맞는 인터넷 강의 선생님을 찾고 싶어. ebsi에서 고등학교 3학년 미적분 강좌를 검색 후 가장 위에 나오는 강좌를 찜해줘.", "step_idx": 1, "raw_step": 2, "image": "images/0714_ebsi_1/frame_3.png", "image_size": [1280, 826], "action": {"operation": "CLICK", "value": null, "target_text": "닫기"}, "element": {"tag": "BUTTON", "text": "닫기", "aria_label": null, "role": null, "id": null, "class": "modal_bottom_close modal_close", "href": null}, "gold_bbox": [767.9, 644.2, 824.7, 665.3], "gold_bbox_norm": [0.59992, 0.7799, 0.6443, 0.80545], "gold_point": [796.3, 654.75], "gold_point_norm": [0.62211, 0.79268], "click_point": [812.29, 650.98], "click_point_norm": [0.6346, 0.78811], "has_element": true, "scored": true, "url": "https://www.ebsi.co.kr/ebs/pot/poti/main.ebs", "page_title": "EBSi 국가대표 고교강의", "previous_actions": ["GOTO https://www.ebsi.co.kr/ebs/pot/poti/main.ebs"], "task_type": "search", "login_required": true, "total_steps": 39, "alt_actions": [], "gold_bbox_999": [599, 779, 644, 805], "gold_point_999": [621, 792], "click_point_999": [634, 787], "mc_candidates": {"A": "CLICK — click(x, y) — 화면의 한 점을 클릭한다", "B": "TYPE — type(x, y, \"text\") — 입력칸에 텍스트를 입력한다", "C": "SCROLL — scroll(\"up\"|\"down\"|\"left\"|\"right\") — 페이지를 한 화면 넘긴다", "D": "HOTKEY — press(\"key\", ...) — 키나 조합키를 누른다", "E": "CLEAR — clear(x, y) — 입력칸의 기존 내용을 지운다", "F": "GO_BACK — go_back() — 이전 페이지로 돌아간다", "G": "DRAG — drag(x1, y1, x2, y2) — 한 점에서 다른 점으로 끈다"}, "mc_answers": ["A"], "mc_answer": "A", "mc_n_options": 7, "mc_needs_point": true, "mc_needs_value": false, "task_id": "0714_ebsi_1"}
4
+ {"sample_id": "0714_google_docs_1#2", "date": "0714", "website": "google_docs", "task_name": "google_docs_1", "instruction": "비타민C의 효능에 관한 글을 적고 싶어. 구글 독스에서 비타민C의 효능에 관한 글을 적어주고 pdf로 다운로드 해줘.", "step_idx": 1, "raw_step": 2, "image": "images/0714_google_docs_1/frame_2.png", "image_size": [1280, 826], "action": {"operation": "CLICK", "value": null, "target_text": "이메일 또는 휴대전화"}, "element": {"tag": "INPUT", "text": null, "aria_label": "이메일 또는 휴대전화", "role": null, "id": "identifierId", "class": "whsOnd zHQkBf", "href": null}, "gold_bbox": [677.2, 151.7, 1231.4, 226.2], "gold_bbox_norm": [0.52906, 0.18365, 0.96204, 0.27385], "gold_point": [954.3, 188.95], "gold_point_norm": [0.74555, 0.22875], "click_point": [889.6, 186.58], "click_point_norm": [0.695, 0.22588], "has_element": true, "scored": true, "url": "https://accounts.google.com/v3/signin/identifier?continue=https%3A%2F%2Fdocs.google.com%2F&dsh=S-1306464316%3A1784019949610766&emr=1&followup=https%3A%2F%2Fdocs.google.com%2F&osid=1&passive=1209600&flowName=GlifWebSignIn&flowEntry=ServiceLogin&ifkv=Ac50bxu4zUBnseXWeGwhxf6scCIFMgepP6dAgTttSf-N5tUkLqp3QMfTLehfzmBUlRteYjfx7_6faA", "page_title": "로그인 - Google 계정", "previous_actions": ["GOTO https://accounts.google.com/v3/signin/identifier?continue=https%3A%2F%2Fdocs.google.com%2F&dsh=S-1306464316%3A1784019949610766&emr=1&followup=https%3A%2F%2Fdocs.google.com%2F&osid=1&passive=1209600&flowName=GlifWebSignIn&flowEntry=ServiceLogin&ifkv=Ac50bxu4zUBnseXWeGwhxf6scCIFMgepP6dAgTttSf-N5tUkLqp3QMfTLehfzmBUlRteYjfx7_6faA"], "task_type": "write", "login_required": true, "total_steps": 15, "alt_actions": [], "gold_bbox_999": [529, 183, 961, 274], "gold_point_999": [745, 229], "click_point_999": [694, 226], "mc_candidates": {"A": "CLICK — click(x, y) — 화면의 한 점을 클릭한다", "B": "TYPE — type(x, y, \"text\") — 입력칸에 텍스트를 입력한다", "C": "SCROLL — scroll(\"up\"|\"down\"|\"left\"|\"right\") — 페이지를 한 화면 넘긴다", "D": "HOTKEY — press(\"key\", ...) — 키나 조합키를 누른다", "E": "CLEAR — clear(x, y) — 입력칸의 기존 내용을 지운다", "F": "GO_BACK — go_back() — 이전 페이지로 돌아간다", "G": "DRAG — drag(x1, y1, x2, y2) — 한 점에서 다른 점으로 끈다"}, "mc_answers": ["A"], "mc_answer": "A", "mc_n_options": 7, "mc_needs_point": true, "mc_needs_value": false, "task_id": "0714_google_docs_1"}
5
+ {"sample_id": "0714_instagram_2_1#2", "date": "0714", "website": "instagram_2", "task_name": "instagram_2_1", "instruction": "인스타그램에서 'kimsseungmini' 계정에 오늘 만나서 정말 반가웠다는 내용의 메세지를 2줄로 보내줘.", "step_idx": 1, "raw_step": 2, "image": "images/0714_instagram_2_1/frame_3.png", "image_size": [1280, 826], "action": {"operation": "CLICK", "value": null, "target_text": null}, "element": {"tag": "INPUT", "text": null, "aria_label": null, "role": null, "id": "_r_2_", "class": "x1i10hfl xggy1nq xtpw4lu x1tutvks x1s3xk63 x1s07b3s x1a2a7pz xjbqb8w x1v8p93f x1o3jo1z x16stqrj xv5lvn5 x1ejq31n x18oe1m7 x1sy0etr xstzfhl x972fbf x10w94by x1qhh985 x14e42zd x9f619 xzsf02u x1lliihq x15h3p50 x10emqs4 x1vr9vpq x1iyjqo2 x10d0gm4 x1fhayk4 x16wdlz0 x3cjxhe xe9ewy2 x11lt19s xeuugli xlyipyv x1hcrkkg xfvqz1d x12vv892 x1hu168l xttzon8 x1sfh74k x3fqe8q x185fvkj x1p97g3g xmtqnhx x11ig0mb xgmu6d7 x1quw8ve xx0ingd xdj266r xyiysdx x14vy60q x109j2v6 xp5op4 x1y44fgy xdzva22 xs8nzd4 x1fzehxr xha3pab", "href": null}, "gold_bbox": [259.1, 347.8, 998.9, 402.3], "gold_bbox_norm": [0.20242, 0.42106, 0.7804, 0.48705], "gold_point": [629.0, 375.05], "gold_point_norm": [0.49141, 0.45406], "click_point": [581.3, 385.08], "click_point_norm": [0.45414, 0.4662], "has_element": true, "scored": true, "url": "https://www.instagram.com/", "page_title": "Instagram", "previous_actions": ["GOTO https://www.instagram.com/"], "task_type": "write", "login_required": true, "total_steps": 25, "alt_actions": [], "gold_bbox_999": [202, 421, 780, 487], "gold_point_999": [491, 454], "click_point_999": [454, 466], "mc_candidates": {"A": "CLICK — click(x, y) — 화면의 한 점을 클릭한다", "B": "TYPE — type(x, y, \"text\") — 입력칸에 텍스트를 입력한다", "C": "SCROLL — scroll(\"up\"|\"down\"|\"left\"|\"right\") — 페이지를 한 화면 넘긴다", "D": "HOTKEY — press(\"key\", ...) — 키나 조합키를 누른다", "E": "CLEAR — clear(x, y) — 입력칸의 기존 내용을 지운다", "F": "GO_BACK — go_back() — 이전 페이지로 돌아간다", "G": "DRAG — drag(x1, y1, x2, y2) — 한 점에서 다른 점으로 끈다"}, "mc_answers": ["A"], "mc_answer": "A", "mc_n_options": 7, "mc_needs_point": true, "mc_needs_value": false, "task_id": "0714_instagram_2_1"}
6
+ {"sample_id": "0714_nate_1#2", "date": "0714", "website": "nate", "task_name": "nate_1", "instruction": "여행계획을 짰는데 메일에 저장해두고 싶어. 네이트에서 내게쓰기로 여행계획 메일을 작성 후 전송해줘.제목은 여행계획으로", "step_idx": 1, "raw_step": 2, "image": "images/0714_nate_1/frame_2.png", "image_size": [1280, 826], "action": {"operation": "CLICK", "value": null, "target_text": null}, "element": {"tag": "INPUT", "text": null, "aria_label": null, "role": null, "id": "ID", "class": "input_id", "href": null}, "gold_bbox": [873.9, 283.0, 1109.0, 310.9], "gold_bbox_norm": [0.68273, 0.34261, 0.86641, 0.3764], "gold_point": [991.45, 296.95], "gold_point_norm": [0.77457, 0.3595], "click_point": [1008.92, 294.17], "click_point_norm": [0.78822, 0.35613], "has_element": true, "scored": true, "url": "https://www.nate.com/", "page_title": "네이트", "previous_actions": ["GOTO https://www.nate.com/"], "task_type": "write", "login_required": true, "total_steps": 14, "alt_actions": [], "gold_bbox_999": [682, 342, 866, 376], "gold_point_999": [774, 359], "click_point_999": [787, 356], "mc_candidates": {"A": "CLICK — click(x, y) — 화면의 한 점을 클릭한다", "B": "TYPE — type(x, y, \"text\") — 입력칸에 텍스트를 입력한다", "C": "SCROLL — scroll(\"up\"|\"down\"|\"left\"|\"right\") — 페이지를 한 화면 넘긴다", "D": "HOTKEY — press(\"key\", ...) — 키나 조합키를 누른다", "E": "CLEAR — clear(x, y) — 입력칸의 기존 내용을 지운다", "F": "GO_BACK — go_back() — 이전 페이지로 돌아간다", "G": "DRAG — drag(x1, y1, x2, y2) — 한 점에서 다른 점으로 끈다"}, "mc_answers": ["A"], "mc_answer": "A", "mc_n_options": 7, "mc_needs_point": true, "mc_needs_value": false, "task_id": "0714_nate_1"}
7
+ {"sample_id": "0714_naver_webtoon_1#2", "date": "0714", "website": "naver_webtoon", "task_name": "naver_webtoon_1", "instruction": "네이버 웹툰에서 무협 검색 후 상위 5개 작품 관심등록 하기", "step_idx": 1, "raw_step": 2, "image": "images/0714_naver_webtoon_1/frame_2.png", "image_size": [1280, 781], "action": {"operation": "CLICK", "value": null, "target_text": "로그인"}, "element": {"tag": "SPAN", "text": "로그인", "aria_label": null, "role": null, "id": null, "class": "gnb_txt", "href": null}, "gold_bbox": [1063.4, 17.5, 1101.9, 34.2], "gold_bbox_norm": [0.83078, 0.0224, 0.86086, 0.0438], "gold_point": [1082.65, 25.85], "gold_point_norm": [0.84582, 0.0331], "click_point": [1092.26, 27.51], "click_point_norm": [0.85332, 0.03522], "has_element": true, "scored": true, "url": "https://comic.naver.com/index", "page_title": "네이버 웹툰", "previous_actions": ["GOTO https://comic.naver.com/index"], "task_type": "search", "login_required": true, "total_steps": 27, "alt_actions": [], "gold_bbox_999": [830, 22, 860, 44], "gold_point_999": [845, 33], "click_point_999": [852, 35], "mc_candidates": {"A": "CLICK — click(x, y) — 화면의 한 점을 클릭한다", "B": "TYPE — type(x, y, \"text\") — 입력칸에 텍스트를 입력한다", "C": "SCROLL — scroll(\"up\"|\"down\"|\"left\"|\"right\") — 페이지를 한 화면 넘긴다", "D": "HOTKEY — press(\"key\", ...) — 키나 조합키를 누른다", "E": "CLEAR — clear(x, y) — 입력칸의 기존 내용을 지운다", "F": "GO_BACK — go_back() — 이전 페이지로 돌아간다", "G": "DRAG — drag(x1, y1, x2, y2) — 한 점에서 다른 점으로 끈다"}, "mc_answers": ["A"], "mc_answer": "A", "mc_n_options": 7, "mc_needs_point": true, "mc_needs_value": false, "task_id": "0714_naver_webtoon_1"}
8
+ {"sample_id": "0714_ohou_se_1#3", "date": "0714", "website": "ohou_se", "task_name": "ohou_se_1", "instruction": "거실에서 사용할 러그를 사려고 하는데 어떤걸 골라야할지 모르겠어. 오늘의 집에서 러그를 검색 해주고 방수가 가능한 제품으로 3개 찜해줘.", "step_idx": 2, "raw_step": 3, "image": "images/0714_ohou_se_1/frame_3.png", "image_size": [1280, 826], "action": {"operation": "CLICK", "value": null, "target_text": null}, "element": {"tag": null, "text": null, "aria_label": null, "role": null, "id": null, "class": null, "href": null}, "gold_bbox": [694.0, 580.0, 752.1, 634.0], "gold_bbox_norm": [0.54218, 0.70217, 0.58758, 0.76756], "gold_point": [723.05, 607.0], "gold_point_norm": [0.56488, 0.73487], "click_point": [723.0, 607.0], "click_point_norm": [0.56484, 0.73487], "has_element": true, "scored": true, "url": null, "page_title": null, "previous_actions": ["GOTO https://ohou.se/", "CLICK 로그인"], "task_type": "search", "login_required": true, "total_steps": 76, "alt_actions": [], "gold_bbox_999": [542, 701, 587, 767], "gold_point_999": [564, 734], "click_point_999": [564, 734], "mc_candidates": {"A": "CLICK — click(x, y) — 화면의 한 점을 클릭한다", "B": "TYPE — type(x, y, \"text\") — 입력칸에 텍스트를 입력한다", "C": "SCROLL — scroll(\"up\"|\"down\"|\"left\"|\"right\") — 페이지를 한 화면 넘긴다", "D": "HOTKEY — press(\"key\", ...) — 키나 조합키를 누른다", "E": "CLEAR — clear(x, y) — 입력칸의 기존 내용을 지운다", "F": "GO_BACK — go_back() — 이전 페이지로 돌아간다", "G": "DRAG — drag(x1, y1, x2, y2) — 한 점에서 다른 점으로 끈다"}, "mc_answers": ["A"], "mc_answer": "A", "mc_n_options": 7, "mc_needs_point": true, "mc_needs_value": false, "task_id": "0714_ohou_se_1"}
9
+ {"sample_id": "0714_pexels_1#2", "date": "0714", "website": "pexels", "task_name": "pexels_1", "instruction": "pexels에서 바다 검색 후 비교하고 다운로드 하기", "step_idx": 1, "raw_step": 2, "image": "images/0714_pexels_1/frame_2.png", "image_size": [1280, 781], "action": {"operation": "CLICK", "value": null, "target_text": null}, "element": {"tag": "INPUT", "text": null, "aria_label": null, "role": null, "id": "search", "class": "SearchInput_input__NOK7W", "href": null}, "gold_bbox": [144.5, 265.7, 631.7, 305.8], "gold_bbox_norm": [0.11289, 0.3402, 0.49352, 0.39155], "gold_point": [388.1, 285.75], "gold_point_norm": [0.3032, 0.36588], "click_point": [345.45, 286.73], "click_point_norm": [0.26988, 0.36713], "has_element": true, "scored": true, "url": "https://www.pexels.com/ko-kr/", "page_title": "무료이미지, 저작권 없는 이미지, 무료 이미지 사이트 | Pexels", "previous_actions": ["GOTO https://www.pexels.com/ko-kr/"], "task_type": "search", "login_required": false, "total_steps": 25, "alt_actions": [], "gold_bbox_999": [113, 340, 493, 391], "gold_point_999": [303, 366], "click_point_999": [270, 367], "mc_candidates": {"A": "CLICK — click(x, y) — 화면의 한 점을 클릭한다", "B": "TYPE — type(x, y, \"text\") — 입력칸에 텍스트를 입력한다", "C": "SCROLL — scroll(\"up\"|\"down\"|\"left\"|\"right\") — 페이지를 한 화면 넘긴다", "D": "HOTKEY — press(\"key\", ...) — 키나 조합키를 누른다", "E": "CLEAR — clear(x, y) — 입력칸의 기존 내용을 지운다", "F": "GO_BACK — go_back() — 이전 페이지로 돌아간다", "G": "DRAG — drag(x1, y1, x2, y2) — 한 점에서 다른 점으로 끈다"}, "mc_answers": ["A"], "mc_answer": "A", "mc_n_options": 7, "mc_needs_point": true, "mc_needs_value": false, "task_id": "0714_pexels_1"}
10
+ {"sample_id": "0714_pinterest_1#2", "date": "0714", "website": "pinterest", "task_name": "pinterest_1", "instruction": "귀여운 고양이 사진을 다운로드 하고싶어. 핀터레스트에서 고양이를 검색하고 가장 첫 번째에 나오는 사진을 다운로드 해줘.", "step_idx": 1, "raw_step": 2, "image": "images/0714_pinterest_1/frame_2.png", "image_size": [1280, 826], "action": {"operation": "CLICK", "value": null, "target_text": "로그인"}, "element": {"tag": "DIV", "text": "로그인", "aria_label": null, "role": null, "id": null, "class": "lIkAnG eMU5i5 o5UlW_ hL1e7w", "href": null}, "gold_bbox": [984.5, 22.9, 1093.6, 91.7], "gold_bbox_norm": [0.76914, 0.02772, 0.85438, 0.11102], "gold_point": [1039.05, 57.3], "gold_point_norm": [0.81176, 0.06937], "click_point": [1043.28, 56.78], "click_point_norm": [0.81506, 0.06875], "has_element": true, "scored": true, "url": "https://kr.pinterest.com/", "page_title": "Pinterest - 핀터레스트", "previous_actions": ["GOTO https://kr.pinterest.com/"], "task_type": "search", "login_required": true, "total_steps": 17, "alt_actions": [], "gold_bbox_999": [768, 28, 854, 111], "gold_point_999": [811, 69], "click_point_999": [814, 69], "mc_candidates": {"A": "CLICK — click(x, y) — 화면의 한 점을 클릭한다", "B": "TYPE — type(x, y, \"text\") — 입력칸에 텍스트를 입력한다", "C": "SCROLL — scroll(\"up\"|\"down\"|\"left\"|\"right\") — 페이지를 한 화면 넘긴다", "D": "HOTKEY — press(\"key\", ...) — 키나 조합키를 누른다", "E": "CLEAR — clear(x, y) — 입력칸의 기존 내용을 지운다", "F": "GO_BACK — go_back() — 이전 페이지로 돌아간다", "G": "DRAG — drag(x1, y1, x2, y2) — 한 점에서 다른 점으로 끈다"}, "mc_answers": ["A"], "mc_answer": "A", "mc_n_options": 7, "mc_needs_point": true, "mc_needs_value": false, "task_id": "0714_pinterest_1"}
examples/test.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4be349b4339c3a060ea0ac90209b6676b0a65efd587c2167fce8602527803cd3
3
+ size 6078994
multimodal_mind2web_ko.tar.gz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85f19a383c5d5d6ca8bd77c0b52fa74295229dde0a7635e97569d375a3d9b172
3
+ size 5026796446
statistics.json ADDED
@@ -0,0 +1,228 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "benchmark": "Ko-MultimodalMind2Web",
3
+ "language": "ko",
4
+ "construction_effort": {
5
+ "korean_human_annotators": 2,
6
+ "hours_per_annotator_per_workday": 4,
7
+ "workdays_per_week": 5,
8
+ "reported_duration_months": 3,
9
+ "approximate_duration_weeks_used_for_estimate": 12,
10
+ "estimated_annotator_hours": 480,
11
+ "estimated_hours_per_annotator": 240,
12
+ "scope": "shared manual collection and annotation project underlying the three related releases",
13
+ "estimate_formula": "2 annotators * 4 hours/day * 5 days/week * approximately 12 weeks"
14
+ },
15
+ "statistics": {
16
+ "scored_steps": 12905,
17
+ "all_scored_flags": true,
18
+ "tasks": 549,
19
+ "rewritten_tasks": 549,
20
+ "unique_images": 12905,
21
+ "websites": 144,
22
+ "operation_distribution": {
23
+ "CLICK": 5879,
24
+ "TYPE": 1399,
25
+ "HOTKEY": 444,
26
+ "SCROLL": 5007,
27
+ "CLEAR": 15,
28
+ "GO_BACK": 134,
29
+ "DRAG": 27
30
+ },
31
+ "primary_target_boxes": 7197,
32
+ "steps_with_primary_or_alternative_boxes": 7580,
33
+ "steps_with_alternative_actions": 755,
34
+ "alternative_actions": 756,
35
+ "alternative_action_sources": {
36
+ "auto:submit-form": 372,
37
+ "manual": 350,
38
+ "auto:template-match": 34
39
+ },
40
+ "scored_steps_per_task": {
41
+ "minimum": 3,
42
+ "median": 21,
43
+ "mean": 23.51,
44
+ "maximum": 100
45
+ },
46
+ "recorded_steps_in_task_metadata": 14042,
47
+ "task_types": {
48
+ "write": 39,
49
+ "login": 243,
50
+ "search": 267
51
+ },
52
+ "by_website": {
53
+ "11st": 79,
54
+ "Hmall": 73,
55
+ "X_2": 9,
56
+ "ably": 136,
57
+ "academyinfo": 51,
58
+ "airport": 256,
59
+ "aladin": 618,
60
+ "amoremall": 81,
61
+ "astro": 77,
62
+ "auction": 129,
63
+ "baskinrobbins": 44,
64
+ "bobaedream": 115,
65
+ "brunch": 54,
66
+ "bunjang": 196,
67
+ "career": 127,
68
+ "cgv": 20,
69
+ "chzzk_naver": 33,
70
+ "cine21": 14,
71
+ "cjlogistics": 7,
72
+ "cjmarket": 105,
73
+ "cjonstyle": 129,
74
+ "comic_naver": 30,
75
+ "compuzone": 115,
76
+ "crocs": 57,
77
+ "daangn": 40,
78
+ "dabang": 111,
79
+ "daiso_mall": 26,
80
+ "dict_naver": 9,
81
+ "discord": 38,
82
+ "duolingo": 199,
83
+ "ebsi": 37,
84
+ "ediya": 76,
85
+ "elevenst": 32,
86
+ "encar": 112,
87
+ "encykorea": 73,
88
+ "epeople": 58,
89
+ "etland": 113,
90
+ "everland": 28,
91
+ "finance_naver": 50,
92
+ "foodsafety": 54,
93
+ "form_naver": 60,
94
+ "gamejob": 117,
95
+ "gamemeca": 107,
96
+ "gcalendar": 143,
97
+ "genie": 98,
98
+ "gmarket": 65,
99
+ "goodwear": 134,
100
+ "google_arts": 90,
101
+ "google_docs": 13,
102
+ "google_maps": 15,
103
+ "googlekeep": 129,
104
+ "hanssem": 95,
105
+ "himart": 559,
106
+ "historydb": 59,
107
+ "historynet": 88,
108
+ "hmall": 151,
109
+ "hogang_nono": 18,
110
+ "hwahae": 118,
111
+ "idus": 346,
112
+ "ikea": 128,
113
+ "inflearn": 38,
114
+ "instagram_2": 115,
115
+ "inven": 87,
116
+ "iris": 140,
117
+ "jobplanet": 141,
118
+ "juso": 28,
119
+ "kakao_webtoon": 157,
120
+ "kakaomap": 174,
121
+ "kbo": 7,
122
+ "kbs": 69,
123
+ "kfa": 89,
124
+ "kin_naver": 111,
125
+ "kipris": 96,
126
+ "kleague": 70,
127
+ "kmdb": 97,
128
+ "kogl": 80,
129
+ "korail": 24,
130
+ "korea_policy": 80,
131
+ "kosaf": 90,
132
+ "kream": 74,
133
+ "kurly": 55,
134
+ "kyobobook": 106,
135
+ "laftel": 99,
136
+ "lezhin": 150,
137
+ "lge": 139,
138
+ "livart": 155,
139
+ "lotte_cinema": 28,
140
+ "lotteon": 119,
141
+ "map_naver": 26,
142
+ "mcdonalds": 16,
143
+ "mega_coffee": 41,
144
+ "megastudy": 8,
145
+ "millie": 107,
146
+ "miricanvas": 19,
147
+ "munpia": 82,
148
+ "nate": 12,
149
+ "national_library": 103,
150
+ "national_science": 58,
151
+ "naver_band": 94,
152
+ "naver_cafe": 8,
153
+ "naver_sports": 11,
154
+ "naver_webtoon": 46,
155
+ "naver_wordwrong": 12,
156
+ "naverband": 18,
157
+ "naverform": 176,
158
+ "navermemo": 132,
159
+ "naverpcgame": 58,
160
+ "nexon": 87,
161
+ "nibr": 71,
162
+ "nolticket": 113,
163
+ "ohou_se": 102,
164
+ "orbi": 138,
165
+ "paikdabang": 62,
166
+ "pann_nate": 58,
167
+ "papago": 8,
168
+ "paris": 33,
169
+ "pexels": 39,
170
+ "pinterest": 30,
171
+ "poom": 59,
172
+ "postype": 96,
173
+ "samsung": 108,
174
+ "saramin": 34,
175
+ "scienceall": 91,
176
+ "sejong": 114,
177
+ "soop": 17,
178
+ "spao": 100,
179
+ "ssfshop": 105,
180
+ "stdict": 103,
181
+ "subway": 5,
182
+ "tistory": 90,
183
+ "todayhouse": 121,
184
+ "tprice": 73,
185
+ "tumblbug": 110,
186
+ "tving": 17,
187
+ "tworld": 25,
188
+ "visitseoul": 89,
189
+ "wadiz": 126,
190
+ "wanted": 196,
191
+ "watchapedia": 77,
192
+ "ypbooks": 98,
193
+ "zara": 42,
194
+ "zigbang": 13,
195
+ "zigzag": 347,
196
+ "zoom": 119
197
+ },
198
+ "instruction_variants_changed": 497
199
+ },
200
+ "gold_label_scorer_sanity": {
201
+ "element_accuracy": {
202
+ "correct": 7580,
203
+ "total": 7580,
204
+ "score": 1.0,
205
+ "credited_without_coordinates": 383
206
+ },
207
+ "step_success_rate": {
208
+ "correct": 12905,
209
+ "total": 12905,
210
+ "score": 1.0
211
+ },
212
+ "task_success_rate": {
213
+ "correct": 549,
214
+ "total": 549,
215
+ "score": 1.0
216
+ }
217
+ },
218
+ "metadata_validation": {
219
+ "rows": 12905,
220
+ "unique_sample_ids": 12905,
221
+ "positive_finite_primary_boxes": 7197,
222
+ "primary_boxes": 7197,
223
+ "out_of_image_primary_boxes": 0,
224
+ "all_image_sizes_positive": true,
225
+ "task_membership_complete": true,
226
+ "task_step_counts_match": true
227
+ }
228
+ }