詳細モード追加
Browse files- .gitignore +5 -1
- app.py +1105 -148
- run.py +1105 -146
.gitignore
CHANGED
@@ -165,4 +165,8 @@ user_annotation/*
|
|
165 |
run.py
|
166 |
backup.py
|
167 |
huggingface_dataset.txt
|
168 |
-
note.txt
|
|
|
|
|
|
|
|
|
|
165 |
run.py
|
166 |
backup.py
|
167 |
huggingface_dataset.txt
|
168 |
+
note.txt
|
169 |
+
type/
|
170 |
+
claude/
|
171 |
+
app_backup.py
|
172 |
+
run copy.py
|
app.py
CHANGED
@@ -26,14 +26,14 @@ else:
|
|
26 |
print("Warning: HF_TOKEN not found. Please set it in your Space secrets.")
|
27 |
|
28 |
# HFデータセット アップロード先
|
29 |
-
# (
|
30 |
-
|
31 |
"team-hatakeyama-phase2/annotation_tanuki_phase2",
|
32 |
"kevineen/Tanuki-Phase2-annotation-dataset", # 出力テスト
|
33 |
]
|
34 |
|
35 |
# アノテーション対象データセット
|
36 |
-
|
37 |
"hatakeyama-llm-team/AutoGeneratedJapaneseQA",
|
38 |
"hatakeyama-llm-team/AutoGeneratedJapaneseQA-other",
|
39 |
"kanhatakeyama/ChatbotArenaJaMixtral8x22b",
|
@@ -44,9 +44,11 @@ annotation_dataset_list = [
|
|
44 |
# "hatakeyama-llm-team/WikiBookJa",
|
45 |
# "kanhatakeyama/AutoWikiQA",
|
46 |
# "susumuota/SyntheticTextWikiTranslate-askllm-v1", # Ask-LLM 翻訳
|
|
|
|
|
47 |
]
|
48 |
|
49 |
-
|
50 |
# マルチターン 未対応
|
51 |
"kanhatakeyama/AutoMultiTurnByMixtral8x22b",
|
52 |
]
|
@@ -58,58 +60,151 @@ is_selected_dataset = gr.State(False)
|
|
58 |
is_loaded_dataset = gr.State(False)
|
59 |
|
60 |
# 選択中のデータセットリスト
|
61 |
-
dropdown_dataset_list = gr.State(value=
|
|
|
62 |
# 現在の対象データセット 初期値は"hatakeyama-llm-team/AutoGeneratedJapaneseQA",
|
63 |
select_dropdown_dataset = gr.State(dropdown_dataset_list.value[0])
|
64 |
select_dataset = gr.State(None) # 現在のデータセット
|
65 |
select_dataset_total_len = gr.State(0) # 現在のデータセットの長さ
|
66 |
select_idx = gr.State(0) # 現在のインデックス (ランダムモードにするなら不要?
|
67 |
-
random_mode = gr.State(False)
|
68 |
|
69 |
# 回答者がアノテーションしたデータセット
|
70 |
annotated_dataset = gr.State(
|
71 |
pd.DataFrame({
|
72 |
-
'dataset_name': [],
|
73 |
-
'dataset_id': [],
|
74 |
-
'who': [],
|
75 |
-
'
|
76 |
-
'
|
77 |
-
'
|
78 |
-
'
|
79 |
-
'is_proofreading_1': [],
|
80 |
'answer_text_1': [],
|
81 |
-
'is_proofreading_2': [], #
|
82 |
-
'answer_text_2': [], #
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
83 |
})
|
84 |
)
|
85 |
|
86 |
-
|
87 |
-
initial_answer_text_2 = gr.State("") # 回答2を整形したかチェック用
|
88 |
-
|
89 |
-
is_dataset_loaded = gr.State(False)
|
90 |
|
91 |
you_dataset_id = gr.State(0) # 回答者がアノテーションしているデータのID
|
92 |
dataset_name = gr.State("") # 編集に使用したデータセット名
|
93 |
dataset_id = gr.State(0) # 加工元データセットのindex
|
94 |
who = gr.State("") # アノテーション者名
|
|
|
|
|
|
|
95 |
good = gr.State(False) # 良
|
96 |
bad = gr.State(False) # 悪
|
97 |
-
|
98 |
-
|
|
|
|
|
99 |
is_proofreading_1 = gr.State(False) # 回答1を整形したか_1
|
100 |
answer_text_1 = gr.State("") # answer_1 回答
|
101 |
is_proofreading_2 = gr.State(False) # 回答2を整形したか_2
|
102 |
answer_text_2 = gr.State("") # answer_2 回答
|
103 |
|
104 |
-
#
|
105 |
-
#
|
106 |
-
|
107 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
def dataset_load_fn() -> Tuple[
|
109 |
str,
|
110 |
str,
|
111 |
str,
|
112 |
str,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
113 |
gr.update,
|
114 |
gr.update,
|
115 |
gr.update,
|
@@ -118,9 +213,12 @@ def dataset_load_fn() -> Tuple[
|
|
118 |
gr.update,
|
119 |
gr.update,
|
120 |
gr.update,
|
121 |
-
gr.update
|
|
|
|
|
|
|
122 |
|
123 |
-
is_dataset_loaded.value = False #
|
124 |
|
125 |
select_dataset.value = load_dataset(
|
126 |
select_dropdown_dataset.value
|
@@ -129,6 +227,7 @@ def dataset_load_fn() -> Tuple[
|
|
129 |
# DatasetオブジェクトをPandas DataFrameに変換
|
130 |
df = select_dataset.value["train"].to_pandas()
|
131 |
|
|
|
132 |
# index列を追加し、シャッフル
|
133 |
df = df.reset_index(drop=False) # 元のindexを保持
|
134 |
df = df.sample(frac=1).reset_index(drop=True) # シャッフル
|
@@ -138,7 +237,7 @@ def dataset_load_fn() -> Tuple[
|
|
138 |
select_dataset_total_len.value = len(df) # 長さを取得
|
139 |
is_dataset_loaded.value = True # ロード完了
|
140 |
|
141 |
-
#
|
142 |
initial_answer_text_1.value = df.iloc[select_idx.value]["answer"]
|
143 |
initial_answer_text_2.value = df.iloc[select_idx.value]["answer"]
|
144 |
|
@@ -146,6 +245,39 @@ def dataset_load_fn() -> Tuple[
|
|
146 |
df.iloc[select_idx.value]["answer"], \
|
147 |
df.iloc[select_idx.value]["question"], \
|
148 |
df.iloc[select_idx.value]["answer"], \
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
149 |
gr.update(interactive=True), \
|
150 |
gr.update(interactive=True), \
|
151 |
gr.update(interactive=True), \
|
@@ -164,7 +296,7 @@ annotation_file = Path("user_annotation/") / f"data_{uuid.uuid4()}.json"
|
|
164 |
annotated_folder = annotation_file.parent
|
165 |
|
166 |
scheduler = CommitScheduler(
|
167 |
-
repo_id=
|
168 |
repo_type="dataset",
|
169 |
folder_path=annotated_folder,
|
170 |
path_in_repo="data", # Spaceの場合の保存先フォルダー
|
@@ -172,19 +304,58 @@ scheduler = CommitScheduler(
|
|
172 |
every=5, # 5分毎にアップロード HuggingFAce_Documentの最低推奨値
|
173 |
)
|
174 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
175 |
# CommitScheduler (HFへのデータアップロード
|
176 |
def save_annotation(
|
177 |
dataset_name: str,
|
178 |
dataset_id: int,
|
179 |
who: str,
|
|
|
180 |
good: bool,
|
181 |
bad: bool,
|
182 |
-
unknown_quality: bool,
|
183 |
-
score: int,
|
184 |
is_proofreading_1: bool,
|
185 |
answer_text_1: str,
|
186 |
is_proofreading_2: bool,
|
187 |
-
answer_text_2: str
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
188 |
|
189 |
annotated_dataset.value = pd.concat([
|
190 |
annotated_dataset.value,
|
@@ -192,33 +363,82 @@ def save_annotation(
|
|
192 |
'dataset_name': [dataset_name],
|
193 |
'dataset_id': [dataset_id],
|
194 |
'who': [who],
|
|
|
195 |
'good': [good],
|
196 |
'bad': [bad],
|
197 |
-
'unknown_quality': [unknown_quality],
|
198 |
-
'score': [score],
|
199 |
'is_proofreading_1': [is_proofreading_1],
|
200 |
"answer_text_1": [answer_text_1],
|
201 |
'is_proofreading_2': [is_proofreading_2],
|
202 |
-
'answer_text_2': [answer_text_2]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
203 |
})], ignore_index=True).reset_index(drop=True)
|
204 |
|
205 |
# 書き込み
|
206 |
with scheduler.lock:
|
207 |
with annotation_file.open("a", encoding='utf-8') as f:
|
208 |
data_to_write = {
|
209 |
-
# "id": , CommitScheduler
|
210 |
"datetime": str(datetime.datetime.now().isoformat()),
|
211 |
"dataset_name": dataset_name,
|
212 |
"dataset_id": int(dataset_id),
|
213 |
"who": who,
|
|
|
214 |
"good": good,
|
215 |
"bad": bad,
|
216 |
-
"unknown_quality": unknown_quality, # 新しいデータフィールドを追加
|
217 |
-
"score": score,
|
218 |
"is_proofreading_1": is_proofreading_1,
|
219 |
"answer_text_1": answer_text_1,
|
220 |
"is_proofreading_2": is_proofreading_2,
|
221 |
"answer_text_2": answer_text_2,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
222 |
}
|
223 |
f.write(json.dumps(data_to_write, ensure_ascii=False))
|
224 |
f.write("\n")
|
@@ -230,7 +450,6 @@ def save_annotation(
|
|
230 |
|
231 |
# ユーザー名表示
|
232 |
|
233 |
-
|
234 |
def hello(profile: gr.OAuthProfile | None) -> Tuple[str, str]:
|
235 |
if profile is None:
|
236 |
return "プライベートデータセット取得のためにログインしてください。", who.value
|
@@ -242,30 +461,40 @@ def hello(profile: gr.OAuthProfile | None) -> Tuple[str, str]:
|
|
242 |
theme_ = gr.themes.Default()
|
243 |
|
244 |
# 後のCSSデザイン変更用
|
245 |
-
|
246 |
-
|
247 |
def load_css():
|
248 |
with open("style.css", "r") as file:
|
249 |
css_content = file.read()
|
250 |
return css_content
|
251 |
|
252 |
-
|
253 |
# Gradio 画面 ============================================
|
254 |
with gr.Blocks(theme=theme_, css=load_css()) as demo:
|
255 |
|
256 |
gr.Markdown("# データセット アノテーション for Tanuki (Phase2)")
|
257 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
258 |
|
259 |
with gr.Tab("アノテーション (シングルターン)"):
|
260 |
|
261 |
-
|
262 |
-
|
263 |
-
gr.LoginButton(value="HuggingFace ログイン",
|
264 |
-
|
265 |
-
|
266 |
# ユーザー名
|
267 |
-
gr_profile_name = gr.Markdown()
|
268 |
-
demo.load(hello, inputs=None, outputs=[gr_profile_name, who])
|
269 |
|
270 |
with gr.Row():
|
271 |
|
@@ -288,76 +517,353 @@ with gr.Blocks(theme=theme_, css=load_css()) as demo:
|
|
288 |
gr_data_load_btn = gr.Button("② データセットを読み込む")
|
289 |
|
290 |
with gr.Column() as content_column:
|
|
|
291 |
with gr.Tab("③ シンプル(良・悪)"):
|
292 |
with gr.Column():
|
293 |
with gr.Row(equal_height=True):
|
294 |
-
|
295 |
-
|
296 |
"分からない", interactive=False) # 「分からない」ボタンを追加
|
297 |
-
|
298 |
|
299 |
-
|
300 |
label="質問: ", lines=5, interactive=False)
|
301 |
|
302 |
-
|
303 |
-
label="回答:
|
304 |
lines=20,
|
305 |
-
interactive=
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
306 |
|
307 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
308 |
|
309 |
-
|
310 |
-
label="質問: ", lines=5, interactive=False)
|
311 |
|
312 |
-
|
313 |
-
|
314 |
-
|
315 |
-
gr_score_3 = gr.Button("3: 普通", interactive=False)
|
316 |
-
gr_score_2 = gr.Button("2: 悪い", interactive=False)
|
317 |
-
gr_score_1 = gr.Button("1: 低品質", interactive=False)
|
318 |
|
319 |
-
|
320 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
321 |
|
322 |
# 5段階評価ボタンのクリックイベントを定義
|
323 |
def score_button_clicked(button_value):
|
|
|
324 |
good.value = False
|
325 |
bad.value = False
|
326 |
-
unknown_quality.value = False
|
327 |
-
score.value = button_value
|
328 |
-
|
329 |
-
gr_data_load_btn.click(
|
330 |
-
dataset_load_fn,
|
331 |
-
inputs=None,
|
332 |
-
outputs=[gr_question_text_1,
|
333 |
-
gr_answer_text_1,
|
334 |
-
gr_question_text_3,
|
335 |
-
gr_answer_text_3,
|
336 |
-
good_btn,
|
337 |
-
unknown_btn, # 「分からない」ボタンを追加
|
338 |
-
bad_btn,
|
339 |
-
gr_score_1,
|
340 |
-
gr_score_2,
|
341 |
-
gr_score_3,
|
342 |
-
gr_score_4,
|
343 |
-
gr_score_5,
|
344 |
-
]
|
345 |
-
)
|
346 |
|
|
|
347 |
def update_annotation(
|
348 |
input_ans_1: str = None,
|
349 |
-
|
350 |
is_good: bool = None, # good/bad を表すフラグを追加
|
351 |
is_unknown: bool = None, # 「分からない」を表すフラグを追加
|
|
|
352 |
score_value: int = None # 5段階評価の値、good/badの場合はNone
|
353 |
-
) -> Tuple[
|
|
|
|
|
|
|
|
|
|
|
|
|
354 |
|
355 |
# good/bad/unknown と score の状態を更新
|
356 |
update_evaluation_state(is_good, is_unknown, score_value)
|
357 |
|
358 |
# 変更を検知 (5段階評価の場合も処理するように変更)
|
359 |
-
update_answer_state(input_ans_1,
|
360 |
-
|
361 |
# 表示更新
|
362 |
# indexを進める
|
363 |
select_idx.value += 1
|
@@ -367,19 +873,31 @@ with gr.Blocks(theme=theme_, css=load_css()) as demo:
|
|
367 |
# ループさせるか、エラー処理を行う
|
368 |
if select_idx.value >= len(df):
|
369 |
select_idx.value = 0
|
|
|
|
|
|
|
|
|
370 |
|
371 |
# データセットに追加
|
372 |
-
# 元のindex番号(dataset_id)を指定して保存
|
373 |
save_annotation_data(df)
|
374 |
|
375 |
# Nextデータ初期化
|
376 |
initialize_next_data(df)
|
|
|
|
|
|
|
377 |
|
378 |
return gr.update(value=df.iloc[select_idx.value]["question"]), \
|
379 |
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
380 |
gr.update(value=df.iloc[select_idx.value]["question"]), \
|
381 |
-
gr.update(value=df.iloc[select_idx.value]["answer"])
|
|
|
|
|
|
|
|
|
382 |
|
|
|
383 |
def update_evaluation_state(is_good, is_unknown, score_value):
|
384 |
if score_value is not None: # 5段階評価の場合
|
385 |
good.value = False
|
@@ -391,100 +909,539 @@ with gr.Blocks(theme=theme_, css=load_css()) as demo:
|
|
391 |
bad.value = not is_good if not is_unknown else False
|
392 |
unknown_quality.value = is_unknown # 「分からない」の状態を設定
|
393 |
|
394 |
-
|
|
|
395 |
if input_ans_1 is not None and initial_answer_text_1.value != input_ans_1:
|
396 |
is_proofreading_1.value = True
|
397 |
answer_text_1.value = input_ans_1
|
398 |
else:
|
399 |
answer_text_1.value = ""
|
400 |
|
401 |
-
|
|
|
402 |
is_proofreading_2.value = True
|
403 |
-
answer_text_2.value =
|
404 |
else:
|
405 |
answer_text_2.value = ""
|
406 |
|
|
|
407 |
def save_annotation_data(df):
|
408 |
save_annotation(
|
409 |
select_dropdown_dataset.value,
|
410 |
# datasetIdは元のindex番号を使用
|
411 |
df.iloc[select_idx.value]['index'],
|
412 |
who.value,
|
|
|
413 |
good.value,
|
414 |
bad.value,
|
415 |
-
unknown_quality.value, # 「分からない」の状態を追加
|
416 |
-
score.value,
|
417 |
is_proofreading_1.value,
|
418 |
answer_text_1.value,
|
419 |
is_proofreading_2.value,
|
420 |
-
answer_text_2.value
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
421 |
)
|
422 |
|
|
|
423 |
def initialize_next_data(df):
|
424 |
is_proofreading_1.value = False
|
425 |
is_proofreading_2.value = False
|
426 |
initial_answer_text_1.value = df.iloc[select_idx.value]["answer"]
|
427 |
initial_answer_text_2.value = df.iloc[select_idx.value]["answer"]
|
428 |
-
|
429 |
-
|
430 |
-
|
431 |
-
|
432 |
-
good_btn.click(
|
433 |
-
good_click,
|
434 |
-
inputs=[
|
435 |
-
gr_answer_text_1,
|
436 |
-
gr_answer_text_3
|
437 |
-
],
|
438 |
-
outputs=[gr_question_text_1,
|
439 |
-
gr_answer_text_1,
|
440 |
-
gr_question_text_3,
|
441 |
-
gr_answer_text_3]
|
442 |
-
)
|
443 |
-
|
444 |
-
def unknown_click(input_ans_1, input_ans_3):
|
445 |
good.value = False
|
446 |
bad.value = False
|
447 |
-
|
448 |
-
|
449 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
450 |
unknown_click,
|
451 |
inputs=[
|
452 |
-
|
453 |
-
|
454 |
],
|
455 |
-
outputs=[
|
456 |
-
|
457 |
-
|
458 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
459 |
)
|
460 |
|
461 |
-
|
462 |
-
|
463 |
-
|
464 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
465 |
bad_click,
|
466 |
inputs=[
|
467 |
-
|
468 |
-
|
469 |
],
|
470 |
-
outputs=[
|
471 |
-
|
472 |
-
|
473 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
474 |
)
|
475 |
|
476 |
# 5段階評価ボタンのクリックイベント
|
477 |
-
|
478 |
-
|
479 |
-
|
480 |
-
|
481 |
-
|
482 |
-
|
483 |
-
|
484 |
-
|
485 |
-
|
486 |
-
|
487 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
488 |
# TODO Tab切り替えで、アノテ済みの一覧を表示する
|
489 |
# with gr.Tab("アノテ済みデータセット(管理画面)"):
|
490 |
# タブを切り替えた時にデータ表示を更新する
|
|
|
26 |
print("Warning: HF_TOKEN not found. Please set it in your Space secrets.")
|
27 |
|
28 |
# HFデータセット アップロード先
|
29 |
+
# (切り替えてテストする用に配列)
|
30 |
+
OUTPUT_DATASET = [
|
31 |
"team-hatakeyama-phase2/annotation_tanuki_phase2",
|
32 |
"kevineen/Tanuki-Phase2-annotation-dataset", # 出力テスト
|
33 |
]
|
34 |
|
35 |
# アノテーション対象データセット
|
36 |
+
ANNOTATION_DATASET = [
|
37 |
"hatakeyama-llm-team/AutoGeneratedJapaneseQA",
|
38 |
"hatakeyama-llm-team/AutoGeneratedJapaneseQA-other",
|
39 |
"kanhatakeyama/ChatbotArenaJaMixtral8x22b",
|
|
|
44 |
# "hatakeyama-llm-team/WikiBookJa",
|
45 |
# "kanhatakeyama/AutoWikiQA",
|
46 |
# "susumuota/SyntheticTextWikiTranslate-askllm-v1", # Ask-LLM 翻訳
|
47 |
+
|
48 |
+
# Nomotron データ (将来用
|
49 |
]
|
50 |
|
51 |
+
MULTITURN_ANNOTATION_DATASET = [
|
52 |
# マルチターン 未対応
|
53 |
"kanhatakeyama/AutoMultiTurnByMixtral8x22b",
|
54 |
]
|
|
|
60 |
is_loaded_dataset = gr.State(False)
|
61 |
|
62 |
# 選択中のデータセットリスト
|
63 |
+
dropdown_dataset_list = gr.State(value=ANNOTATION_DATASET)
|
64 |
+
|
65 |
# 現在の対象データセット 初期値は"hatakeyama-llm-team/AutoGeneratedJapaneseQA",
|
66 |
select_dropdown_dataset = gr.State(dropdown_dataset_list.value[0])
|
67 |
select_dataset = gr.State(None) # 現在のデータセット
|
68 |
select_dataset_total_len = gr.State(0) # 現在のデータセットの長さ
|
69 |
select_idx = gr.State(0) # 現在のインデックス (ランダムモードにするなら不要?
|
70 |
+
# random_mode = gr.State(False) # ランダムモード
|
71 |
|
72 |
# 回答者がアノテーションしたデータセット
|
73 |
annotated_dataset = gr.State(
|
74 |
pd.DataFrame({
|
75 |
+
'dataset_name': [], # 対象データセット
|
76 |
+
'dataset_id': [], # データセットindex
|
77 |
+
'who': [], # アノテーション者名
|
78 |
+
'unknown_quality': [], # 不明能
|
79 |
+
'good': [], # 良
|
80 |
+
'bad': [], # 悪
|
81 |
+
'is_proofreading_1': [], # テキストに修正があったか
|
|
|
82 |
'answer_text_1': [],
|
83 |
+
'is_proofreading_2': [], # 2ターン目用
|
84 |
+
'answer_text_2': [], # 2ターン目用
|
85 |
+
|
86 |
+
# 5段階評価
|
87 |
+
'score': [], # 総合 (5段階評価)
|
88 |
+
'helphulness': [], # 有用性 (5段階評価)
|
89 |
+
'correcthness': [], # 正確さ (5段階評価)
|
90 |
+
'coherence': [], # 一貫性 (5段階評価)
|
91 |
+
'complexity': [], # 複雑さ (5段階評価)
|
92 |
+
'verbosity': [], # 冗長性 (5段階評価)
|
93 |
+
'humor': [], # ユーモア (SteerLM)
|
94 |
+
'creativity': [], # 創造性 (SteerLM)
|
95 |
+
'appropriate': [], # 適切性 (SteerLMでは不適)
|
96 |
+
'following_instructions': [], # 指示が合った場合、従っているか (SteerLMのfails_taskに該当)
|
97 |
+
'politeness': [], # 礼儀正しさ (minnadechat)
|
98 |
+
'harmfulness': [], # 有性 (minnadechat)
|
99 |
+
|
100 |
+
# 0: 分からない(未設定) 1: 無 2: 有
|
101 |
+
'typos': [], # 誤字脱字 (minnadechat)
|
102 |
+
'hate': [], # ヘイト
|
103 |
+
'sexual': [], # 性的
|
104 |
+
'violence': [], # 暴力
|
105 |
+
'suicide': [], # 自殺
|
106 |
+
'threat': [], # 犯罪
|
107 |
+
'gun': [], # 重火器
|
108 |
+
'controlled_substance': [], # 規制対象物質
|
109 |
+
'climinal_planing': [], # 犯罪計画
|
110 |
+
'privacy': [], # 個人情報
|
111 |
+
'harassment': [], # ハラスメント
|
112 |
+
'profanity': [], # 冒涜
|
113 |
+
'political_content': [], # 政治的内容 (SteerLM)
|
114 |
+
'moral_judgement': [], # 非道徳 (SteerLM)
|
115 |
})
|
116 |
)
|
117 |
|
118 |
+
is_dataset_loaded = gr.State(False) # データセットロード状態
|
|
|
|
|
|
|
119 |
|
120 |
you_dataset_id = gr.State(0) # 回答者がアノテーションしているデータのID
|
121 |
dataset_name = gr.State("") # 編集に使用したデータセット名
|
122 |
dataset_id = gr.State(0) # 加工元データセットのindex
|
123 |
who = gr.State("") # アノテーション者名
|
124 |
+
|
125 |
+
# シンプルモード
|
126 |
+
unknown_quality = gr.State(False) # 分からない(未設定)
|
127 |
good = gr.State(False) # 良
|
128 |
bad = gr.State(False) # 悪
|
129 |
+
|
130 |
+
initial_answer_text_1 = gr.State("") # 回��1を整形したかチェック用
|
131 |
+
initial_answer_text_2 = gr.State("") # 回答2を整形したかチェック用
|
132 |
+
|
133 |
is_proofreading_1 = gr.State(False) # 回答1を整形したか_1
|
134 |
answer_text_1 = gr.State("") # answer_1 回答
|
135 |
is_proofreading_2 = gr.State(False) # 回答2を整形したか_2
|
136 |
answer_text_2 = gr.State("") # answer_2 回答
|
137 |
|
138 |
+
# (5段階評価)
|
139 |
+
# 0は判断していない、又は判断が難しいという特徴量にする
|
140 |
+
score = gr.State(3) # 総合スコア 初期値は3 (最低限必須)
|
141 |
+
helphulness = gr.State(0) # 有用性
|
142 |
+
correcthness = gr.State(0) # 正確さ
|
143 |
+
coherenece = gr.State(0) # 一貫性
|
144 |
+
complexity = gr.State(0) # 複雑さ
|
145 |
+
verbosity = gr.State(0) # 冗長性
|
146 |
+
humor = gr.State(0) # ユーモア (SteerLM)
|
147 |
+
creativity = gr.State(0) # 創造性 (SteerLM)
|
148 |
+
appropriate = gr.State(0) # 適切性 (SteerLMではnot_appropriate)
|
149 |
+
following_instructions = gr.State(0) # 指示への忠実度 (SteerLMのfails_task)
|
150 |
+
politeness = gr.State(0) # 礼儀正しさ (minnadechat)
|
151 |
+
harmfulness = gr.State(0) # 有度 (minnadechat)
|
152 |
+
|
153 |
+
# (有無評価)
|
154 |
+
# 0は判断していない、又は判断が難しいという特徴量にする
|
155 |
+
hate = gr.State(0) # ヘイト
|
156 |
+
sexual = gr.State(0) # 性的
|
157 |
+
violence = gr.State(0) # 暴力
|
158 |
+
suicide = gr.State(0) # 自殺
|
159 |
+
threat = gr.State(0) # 犯罪
|
160 |
+
gun = gr.State(0) # 銃・重火器
|
161 |
+
controlled_substance = gr.State(0) # 規制対象物質
|
162 |
+
climinal_planing = gr.State(0) # 犯罪計画
|
163 |
+
privacy = gr.State(0) # 個人情報・プライバシー情報
|
164 |
+
harassment = gr.State(0) # ハラスメント
|
165 |
+
profanity = gr.State(0) # 冒涜
|
166 |
+
political_content = gr.State(0) # 政治的内容 (SteerLM)
|
167 |
+
moral_judgement = gr.State(0) # 非道徳 (SteerLM)
|
168 |
+
typos = gr.State(0) # 誤字脱字(minnadechat)
|
169 |
+
|
170 |
+
# 未整理 ========================================
|
171 |
+
|
172 |
+
# データセット読み込み
|
173 |
def dataset_load_fn() -> Tuple[
|
174 |
str,
|
175 |
str,
|
176 |
str,
|
177 |
str,
|
178 |
+
str,
|
179 |
+
str,
|
180 |
+
gr.update,
|
181 |
+
gr.update,
|
182 |
+
gr.update,
|
183 |
+
gr.update,
|
184 |
+
gr.update,
|
185 |
+
gr.update,
|
186 |
+
gr.update,
|
187 |
+
gr.update,
|
188 |
+
gr.update,
|
189 |
+
gr.update,
|
190 |
+
gr.update,
|
191 |
+
gr.update,
|
192 |
+
gr.update,
|
193 |
+
gr.update,
|
194 |
+
gr.update,
|
195 |
+
gr.update,
|
196 |
+
gr.update,
|
197 |
+
gr.update,
|
198 |
+
gr.update,
|
199 |
+
gr.update,
|
200 |
+
gr.update,
|
201 |
+
gr.update,
|
202 |
+
gr.update,
|
203 |
+
gr.update,
|
204 |
+
gr.update,
|
205 |
+
gr.update,
|
206 |
+
gr.update,
|
207 |
+
gr.update,
|
208 |
gr.update,
|
209 |
gr.update,
|
210 |
gr.update,
|
|
|
213 |
gr.update,
|
214 |
gr.update,
|
215 |
gr.update,
|
216 |
+
gr.update,
|
217 |
+
gr.update,
|
218 |
+
gr.update,
|
219 |
+
]:
|
220 |
|
221 |
+
is_dataset_loaded.value = False # ロード状態初期化
|
222 |
|
223 |
select_dataset.value = load_dataset(
|
224 |
select_dropdown_dataset.value
|
|
|
227 |
# DatasetオブジェクトをPandas DataFrameに変換
|
228 |
df = select_dataset.value["train"].to_pandas()
|
229 |
|
230 |
+
# ランダム化実施
|
231 |
# index列を追加し、シャッフル
|
232 |
df = df.reset_index(drop=False) # 元のindexを保持
|
233 |
df = df.sample(frac=1).reset_index(drop=True) # シャッフル
|
|
|
237 |
select_dataset_total_len.value = len(df) # 長さを取得
|
238 |
is_dataset_loaded.value = True # ロード完了
|
239 |
|
240 |
+
# データロード時にテキストの初期値を設定
|
241 |
initial_answer_text_1.value = df.iloc[select_idx.value]["answer"]
|
242 |
initial_answer_text_2.value = df.iloc[select_idx.value]["answer"]
|
243 |
|
|
|
245 |
df.iloc[select_idx.value]["answer"], \
|
246 |
df.iloc[select_idx.value]["question"], \
|
247 |
df.iloc[select_idx.value]["answer"], \
|
248 |
+
df.iloc[select_idx.value]["question"], \
|
249 |
+
df.iloc[select_idx.value]["answer"], \
|
250 |
+
gr.update(interactive=True), \
|
251 |
+
gr.update(interactive=True), \
|
252 |
+
gr.update(interactive=True), \
|
253 |
+
gr.update(interactive=True), \
|
254 |
+
gr.update(interactive=True), \
|
255 |
+
gr.update(interactive=True), \
|
256 |
+
gr.update(interactive=True), \
|
257 |
+
gr.update(interactive=True), \
|
258 |
+
gr.update(interactive=True), \
|
259 |
+
gr.update(interactive=True), \
|
260 |
+
gr.update(interactive=True), \
|
261 |
+
gr.update(interactive=True), \
|
262 |
+
gr.update(interactive=True), \
|
263 |
+
gr.update(interactive=True), \
|
264 |
+
gr.update(interactive=True), \
|
265 |
+
gr.update(interactive=True), \
|
266 |
+
gr.update(interactive=True), \
|
267 |
+
gr.update(interactive=True), \
|
268 |
+
gr.update(interactive=True), \
|
269 |
+
gr.update(interactive=True), \
|
270 |
+
gr.update(interactive=True), \
|
271 |
+
gr.update(interactive=True), \
|
272 |
+
gr.update(interactive=True), \
|
273 |
+
gr.update(interactive=True), \
|
274 |
+
gr.update(interactive=True), \
|
275 |
+
gr.update(interactive=True), \
|
276 |
+
gr.update(interactive=True), \
|
277 |
+
gr.update(interactive=True), \
|
278 |
+
gr.update(interactive=True), \
|
279 |
+
gr.update(interactive=True), \
|
280 |
+
gr.update(interactive=True), \
|
281 |
gr.update(interactive=True), \
|
282 |
gr.update(interactive=True), \
|
283 |
gr.update(interactive=True), \
|
|
|
296 |
annotated_folder = annotation_file.parent
|
297 |
|
298 |
scheduler = CommitScheduler(
|
299 |
+
repo_id=OUTPUT_DATASET[1],
|
300 |
repo_type="dataset",
|
301 |
folder_path=annotated_folder,
|
302 |
path_in_repo="data", # Spaceの場合の保存先フォルダー
|
|
|
304 |
every=5, # 5分毎にアップロード HuggingFAce_Documentの最低推奨値
|
305 |
)
|
306 |
|
307 |
+
# チェックボックスをintに変換
|
308 |
+
def checkbox_to_int(checkbox_value) -> int:
|
309 |
+
if checkbox_value == "不明":
|
310 |
+
output = 0
|
311 |
+
elif checkbox_value == "有":
|
312 |
+
output = 1
|
313 |
+
elif checkbox_value == "無":
|
314 |
+
output = 2
|
315 |
+
else :
|
316 |
+
output = 0
|
317 |
+
print("error: ")
|
318 |
+
return output
|
319 |
+
|
320 |
# CommitScheduler (HFへのデータアップロード
|
321 |
def save_annotation(
|
322 |
dataset_name: str,
|
323 |
dataset_id: int,
|
324 |
who: str,
|
325 |
+
unknown_quality: bool,
|
326 |
good: bool,
|
327 |
bad: bool,
|
|
|
|
|
328 |
is_proofreading_1: bool,
|
329 |
answer_text_1: str,
|
330 |
is_proofreading_2: bool,
|
331 |
+
answer_text_2: str,
|
332 |
+
score: int,
|
333 |
+
helphulness: int,
|
334 |
+
correcthness: int,
|
335 |
+
coherence: int,
|
336 |
+
complexity: int,
|
337 |
+
verbosity: int,
|
338 |
+
humor: int,
|
339 |
+
creativity: int,
|
340 |
+
appropriate: int,
|
341 |
+
following_instructions: int,
|
342 |
+
politeness: int,
|
343 |
+
harmfulness: int,
|
344 |
+
hate: int,
|
345 |
+
sexual: int,
|
346 |
+
violence: int,
|
347 |
+
suicide: int,
|
348 |
+
threat: int,
|
349 |
+
gun: int,
|
350 |
+
controlled_substance: int,
|
351 |
+
climinal_planing: int,
|
352 |
+
privacy: int,
|
353 |
+
harassment: int,
|
354 |
+
profanity: int,
|
355 |
+
political_content: int,
|
356 |
+
moral_judgement: int,
|
357 |
+
typos: int,
|
358 |
+
) -> None:
|
359 |
|
360 |
annotated_dataset.value = pd.concat([
|
361 |
annotated_dataset.value,
|
|
|
363 |
'dataset_name': [dataset_name],
|
364 |
'dataset_id': [dataset_id],
|
365 |
'who': [who],
|
366 |
+
'unknown_quality': [unknown_quality],
|
367 |
'good': [good],
|
368 |
'bad': [bad],
|
|
|
|
|
369 |
'is_proofreading_1': [is_proofreading_1],
|
370 |
"answer_text_1": [answer_text_1],
|
371 |
'is_proofreading_2': [is_proofreading_2],
|
372 |
+
'answer_text_2': [answer_text_2],
|
373 |
+
'score': [score],
|
374 |
+
'helphullness': [helphulness],
|
375 |
+
'correcthness': [correcthness],
|
376 |
+
'complexity': [complexity],
|
377 |
+
'verbosity': [verbosity],
|
378 |
+
'humor': [humor],
|
379 |
+
'creativity': [creativity],
|
380 |
+
'appropriate': [appropriate],
|
381 |
+
'following_finstructions': [following_instructions],
|
382 |
+
'politeness': [politeness],
|
383 |
+
'harmfulness': [harmfulness],
|
384 |
+
'hate': [hate],
|
385 |
+
'sextual': [sexual],
|
386 |
+
'violence': [violence],
|
387 |
+
'suicide': [suicide],
|
388 |
+
'threat': [threat],
|
389 |
+
'gun': [gun],
|
390 |
+
'controlled_substance': [controlled_substance],
|
391 |
+
'climinal_planing': [climinal_planing],
|
392 |
+
'privacy': [privacy],
|
393 |
+
'harassment': [harassment],
|
394 |
+
'profanity': [profanity],
|
395 |
+
'political_content': [political_content],
|
396 |
+
'moral_judgement': [moral_judgement],
|
397 |
+
'types': [typos]
|
398 |
})], ignore_index=True).reset_index(drop=True)
|
399 |
|
400 |
# 書き込み
|
401 |
with scheduler.lock:
|
402 |
with annotation_file.open("a", encoding='utf-8') as f:
|
403 |
data_to_write = {
|
404 |
+
# "id": , CommitSchedulerは、取得して末尾idを付与することが不可能?
|
405 |
"datetime": str(datetime.datetime.now().isoformat()),
|
406 |
"dataset_name": dataset_name,
|
407 |
"dataset_id": int(dataset_id),
|
408 |
"who": who,
|
409 |
+
"unknown_quality": unknown_quality,
|
410 |
"good": good,
|
411 |
"bad": bad,
|
|
|
|
|
412 |
"is_proofreading_1": is_proofreading_1,
|
413 |
"answer_text_1": answer_text_1,
|
414 |
"is_proofreading_2": is_proofreading_2,
|
415 |
"answer_text_2": answer_text_2,
|
416 |
+
"score": int(score),
|
417 |
+
"helphulness": int(helphulness),
|
418 |
+
"correcthness": int(correcthness),
|
419 |
+
"coherence": int(coherence),
|
420 |
+
"complexity": int(complexity),
|
421 |
+
"verbosity": int(verbosity),
|
422 |
+
"humor": int(humor),
|
423 |
+
"creativity": int(creativity),
|
424 |
+
"appropriate": int(appropriate),
|
425 |
+
"following_instructions": int(following_instructions),
|
426 |
+
"politeness": int(politeness),
|
427 |
+
"harmfulness": int(harmfulness),
|
428 |
+
"hate": int(hate),
|
429 |
+
"sexual": int(sexual),
|
430 |
+
"violence": int(violence),
|
431 |
+
"suicide": int(suicide),
|
432 |
+
"threat": int(threat),
|
433 |
+
"gun": int(gun),
|
434 |
+
"controlled_substance": int(controlled_substance),
|
435 |
+
"climinal_planing": int(climinal_planing),
|
436 |
+
"privacy": int(privacy),
|
437 |
+
"harassment": int(harassment),
|
438 |
+
"profanity": int(profanity),
|
439 |
+
"policital_content": int(political_content),
|
440 |
+
"moral_judgement": int(moral_judgement),
|
441 |
+
"typos": int(typos)
|
442 |
}
|
443 |
f.write(json.dumps(data_to_write, ensure_ascii=False))
|
444 |
f.write("\n")
|
|
|
450 |
|
451 |
# ユーザー名表示
|
452 |
|
|
|
453 |
def hello(profile: gr.OAuthProfile | None) -> Tuple[str, str]:
|
454 |
if profile is None:
|
455 |
return "プライベートデータセット取得のためにログインしてください。", who.value
|
|
|
461 |
theme_ = gr.themes.Default()
|
462 |
|
463 |
# 後のCSSデザイン変更用
|
|
|
|
|
464 |
def load_css():
|
465 |
with open("style.css", "r") as file:
|
466 |
css_content = file.read()
|
467 |
return css_content
|
468 |
|
|
|
469 |
# Gradio 画面 ============================================
|
470 |
with gr.Blocks(theme=theme_, css=load_css()) as demo:
|
471 |
|
472 |
gr.Markdown("# データセット アノテーション for Tanuki (Phase2)")
|
473 |
+
|
474 |
+
with gr.Row():
|
475 |
+
gr.Markdown("### GENIACにて開発中のLLM用データセットに対してアノテーションするSpaceです\n \
|
476 |
+
入力されたデータは使用される可能性があるため、個人情報・秘匿情報などは入力しないでください。\n \
|
477 |
+
テスト中です。")
|
478 |
+
gr_who = gr.TextArea(value=who.value, lines=1, max_lines=1, label="ユーザー名 (入力してください)")
|
479 |
+
|
480 |
+
def change_name(name: str):
|
481 |
+
who.value = name
|
482 |
+
|
483 |
+
gr_who.change(
|
484 |
+
change_name,
|
485 |
+
inputs=[gr_who],
|
486 |
+
outputs=[]
|
487 |
+
)
|
488 |
|
489 |
with gr.Tab("アノテーション (シングルターン)"):
|
490 |
|
491 |
+
# HF login 機能
|
492 |
+
# with gr.Row(equal_height=True):
|
493 |
+
# gr.LoginButton(value="HuggingFace ログイン",
|
494 |
+
# logout_value="HuggingFace ログアウト", scale=1)
|
|
|
495 |
# ユーザー名
|
496 |
+
# gr_profile_name = gr.Markdown()
|
497 |
+
# demo.load(hello, inputs=None, outputs=[gr_profile_name, who])
|
498 |
|
499 |
with gr.Row():
|
500 |
|
|
|
517 |
gr_data_load_btn = gr.Button("② データセットを読み込む")
|
518 |
|
519 |
with gr.Column() as content_column:
|
520 |
+
|
521 |
with gr.Tab("③ シンプル(良・悪)"):
|
522 |
with gr.Column():
|
523 |
with gr.Row(equal_height=True):
|
524 |
+
gr_good_btn = gr.Button("良い", interactive=False)
|
525 |
+
gr_unknown_btn = gr.Button(
|
526 |
"分からない", interactive=False) # 「分からない」ボタンを追加
|
527 |
+
gr_bad_btn = gr.Button("悪い", interactive=False)
|
528 |
|
529 |
+
gr_question_text_1_1 = gr.Textbox(
|
530 |
label="質問: ", lines=5, interactive=False)
|
531 |
|
532 |
+
gr_answer_text_1_1 = gr.Textbox(
|
533 |
+
label="回答: 訂正して頂けると、品質が上がります。",
|
534 |
lines=20,
|
535 |
+
interactive=False)
|
536 |
+
|
537 |
+
with gr.Tab("③ 5段階評価(シンプル)"):
|
538 |
+
|
539 |
+
with gr.Row() as simple_score_btn:
|
540 |
+
gr_score_5_btn = gr.Button("5: 高品質", interactive=False)
|
541 |
+
gr_score_4_btn = gr.Button("4: 良い", interactive=False)
|
542 |
+
gr_score_3_btn = gr.Button("3: 普通", interactive=False)
|
543 |
+
gr_score_2_btn = gr.Button("2: 悪い", interactive=False)
|
544 |
+
gr_score_1_btn = gr.Button("1: 低品質", interactive=False)
|
545 |
+
|
546 |
+
gr_question_text_2_1 = gr.Textbox(
|
547 |
+
label="質問: ", lines=5, interactive=False)
|
548 |
|
549 |
+
gr_answer_text_2_1 = gr.Textbox(
|
550 |
+
label="回答: 訂正して頂けると、品質が上がります。", lines=20, interactive=False)
|
551 |
+
|
552 |
+
with gr.Tab("③ 5段階評価 (詳細)"):
|
553 |
+
|
554 |
+
with gr.Row():
|
555 |
+
|
556 |
+
with gr.Column() as EvalFive:
|
557 |
+
|
558 |
+
gr_question_text_3_1 = gr.Textbox(
|
559 |
+
label="質問: ", lines=10, interactive=False)
|
560 |
+
|
561 |
+
# gr_explain = gr.Markdown("回答")
|
562 |
+
|
563 |
+
gr_answer_text_3_1 = gr.Textbox(
|
564 |
+
label="回答: 訂正して頂けると、品質が上がります。", lines=35, interactive=False)
|
565 |
+
|
566 |
+
with gr.Column() as EvalFiveDetail:
|
567 |
+
|
568 |
+
with gr.Row():
|
569 |
+
gr_submit_score = gr.Button("評価送信", interactive=False)
|
570 |
+
gr_score_reset = gr.Button("スコアリセット", interactive=False)
|
571 |
+
|
572 |
+
gr_score_detail = gr.Slider(label="総合スコア 【必須】",value=3, minimum=1, maximum=5, step=1, interactive=False)
|
573 |
+
gr_evalannotation_explain = gr.Markdown("詳細アノテーション (5点満点)")
|
574 |
+
|
575 |
+
with gr.Row() :
|
576 |
+
gr_helphulness = gr.Slider(label="有用性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
577 |
+
gr_correcthness = gr.Slider(label="正確さ", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
578 |
+
gr_coherence = gr.Slider(label="一貫性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
579 |
+
gr_complexity = gr.Slider(label="複雑さ", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
580 |
+
|
581 |
+
with gr.Row() :
|
582 |
+
gr_verbosity = gr.Slider(label="冗長性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
583 |
+
gr_humor = gr.Slider(label="ユーモア", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
584 |
+
gr_creativity = gr.Slider(label="創造性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
585 |
+
gr_appropriate = gr.Slider(label="適切性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
586 |
+
|
587 |
+
with gr.Row() :
|
588 |
+
gr_following_instructions = gr.Slider(label="忠実性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
589 |
+
gr_politeness = gr.Slider(label="礼儀正しさ", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
590 |
+
gr_harmfulness = gr.Slider(label="有害性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
591 |
+
gr_text_4 = gr.Markdown("判断可能な物のみ\n\nスコアを付けてください。")
|
592 |
+
|
593 |
+
gr_text_5 = gr.Markdown("")
|
594 |
+
|
595 |
+
with gr.Row():
|
596 |
+
gr_typos = gr.Radio(label="誤字・脱字", choices=["不明", "有", "無"], value="不明", interactive=False)
|
597 |
+
gr_moral_judgement = gr.Radio(label="非道徳", choices=["不明", "有", "無"], value="不明", interactive=False)
|
598 |
+
|
599 |
+
with gr.Row():
|
600 |
+
gr_hate = gr.Radio(label="ヘイト", choices=["不明", "有", "無"], value="不明", interactive=False)
|
601 |
+
gr_sexual = gr.Radio(label="性的内容", choices=["不明", "有", "無"], value="不明", interactive=False)
|
602 |
+
|
603 |
+
with gr.Row():
|
604 |
+
gr_violence = gr.Radio(label="暴力的", choices=["不明", "有", "無"], value="不��", interactive=False)
|
605 |
+
gr_suicide = gr.Radio(label="自殺行為", choices=["不明", "有", "無"], value="不明", interactive=False)
|
606 |
+
|
607 |
+
with gr.Row():
|
608 |
+
gr_threat = gr.Radio(label="犯罪", choices=["不明", "有", "無"], value="不明", interactive=False)
|
609 |
+
gr_gun = gr.Radio(label="銃等", choices=["不明", "有", "無"], value="不明", interactive=False)
|
610 |
+
|
611 |
+
with gr.Row():
|
612 |
+
gr_controlled_substance = gr.Radio(label="規制対象物質", choices=["不明", "有", "無"], value="不明", interactive=False)
|
613 |
+
gr_climinal_planing = gr.Radio(label="犯罪計画", choices=["不明", "有", "無"], value="不明", interactive=False)
|
614 |
+
|
615 |
+
with gr.Row():
|
616 |
+
gr_privacy = gr.Radio(label="個人情報", choices=["不明", "有", "無"], value="不明", interactive=False)
|
617 |
+
gr_harassment = gr.Radio(label="ハラスメント", choices=["不明", "有", "無"], value="不明", interactive=False)
|
618 |
+
|
619 |
+
with gr.Row():
|
620 |
+
gr_profanity = gr.Radio(label="冒涜行為", choices=["不明", "有", "無"], value="不明", interactive=False)
|
621 |
+
gr_political_content = gr.Radio(label="政治的内容", choices=["不明", "有", "無"], value="不明", interactive=False)
|
622 |
+
|
623 |
+
|
624 |
+
# 5段階評価 詳細のgr.State Update
|
625 |
+
def eval_submit(
|
626 |
+
# text
|
627 |
+
g_answer_text_3_1: str,
|
628 |
+
|
629 |
+
# slicer
|
630 |
+
g_score: int,
|
631 |
+
g_helphulness: int,
|
632 |
+
g_correcthness: int,
|
633 |
+
g_coherenece: int,
|
634 |
+
g_complexity: int,
|
635 |
+
g_verbosity: int,
|
636 |
+
g_humor: int,
|
637 |
+
g_creativity: int,
|
638 |
+
g_appropriate: int,
|
639 |
+
g_following_instructions: int,
|
640 |
+
g_politeness: int,
|
641 |
+
g_harmfulness: int,
|
642 |
+
|
643 |
+
# checkbox
|
644 |
+
g_hate: str,
|
645 |
+
g_sexual: str,
|
646 |
+
g_violence: str,
|
647 |
+
g_suicide: str,
|
648 |
+
g_threat: str,
|
649 |
+
g_gun: str,
|
650 |
+
g_controlled_substance: str,
|
651 |
+
g_climinal_planing: str,
|
652 |
+
g_privacy: str,
|
653 |
+
g_harassment: str,
|
654 |
+
g_profanity: str,
|
655 |
+
g_political_content: str,
|
656 |
+
g_moral_judgement: str,
|
657 |
+
g_typos: str,
|
658 |
+
):
|
659 |
+
# 状態初期化
|
660 |
+
good.value = False
|
661 |
+
bad.value = False
|
662 |
+
unknown_quality.value = False
|
663 |
+
|
664 |
+
# テキストに更新があったかどうかチェック
|
665 |
+
if initial_answer_text_1.value != g_answer_text_3_1:
|
666 |
+
is_proofreading_1.value = True
|
667 |
+
answer_text_1.value = g_answer_text_3_1
|
668 |
+
else:
|
669 |
+
is_proofreading_1.value = False
|
670 |
+
answer_text_1.value = ""
|
671 |
+
|
672 |
+
# 送信用データ更新
|
673 |
+
score.value = g_score
|
674 |
+
helphulness.value = g_helphulness
|
675 |
+
correcthness.value = g_correcthness
|
676 |
+
coherenece.value = g_coherenece
|
677 |
+
complexity.value = g_complexity
|
678 |
+
verbosity.value = g_verbosity
|
679 |
+
humor.value = g_humor
|
680 |
+
creativity.value = g_creativity
|
681 |
+
appropriate.value = g_appropriate
|
682 |
+
following_instructions.value = g_following_instructions
|
683 |
+
politeness.value = g_politeness
|
684 |
+
harmfulness.value = g_harmfulness
|
685 |
+
|
686 |
+
hate.value = checkbox_to_int(g_hate)
|
687 |
+
sexual.value = checkbox_to_int(g_sexual)
|
688 |
+
violence.value = checkbox_to_int(g_violence)
|
689 |
+
suicide.value = checkbox_to_int(g_suicide)
|
690 |
+
threat.value = checkbox_to_int(g_threat)
|
691 |
+
gun.value = checkbox_to_int(g_gun)
|
692 |
+
controlled_substance.value = checkbox_to_int(g_controlled_substance)
|
693 |
+
climinal_planing.value = checkbox_to_int(g_climinal_planing)
|
694 |
+
privacy.value = checkbox_to_int(g_privacy)
|
695 |
+
harassment.value = checkbox_to_int(g_harassment)
|
696 |
+
profanity.value = checkbox_to_int(g_profanity)
|
697 |
+
political_content.value = checkbox_to_int(g_political_content)
|
698 |
+
moral_judgement.value = checkbox_to_int(g_moral_judgement)
|
699 |
+
typos.value = checkbox_to_int(g_typos)
|
700 |
+
|
701 |
+
# データ送信
|
702 |
+
|
703 |
+
# 表示更新
|
704 |
+
# indexを進める
|
705 |
+
select_idx.value += 1
|
706 |
|
707 |
+
df = select_dataset.value["train"]
|
|
|
708 |
|
709 |
+
# ループさせるか、エラー処理を行う
|
710 |
+
if select_idx.value >= len(df):
|
711 |
+
select_idx.value = 0
|
|
|
|
|
|
|
712 |
|
713 |
+
# データセットに追加
|
714 |
+
df = select_dataset.value["train"]
|
715 |
+
save_annotation(
|
716 |
+
select_dropdown_dataset.value,
|
717 |
+
df.iloc[select_idx.value]['index'],
|
718 |
+
who.value,
|
719 |
+
unknown_quality.value,
|
720 |
+
good.value,
|
721 |
+
bad.value,
|
722 |
+
is_proofreading_1.value,
|
723 |
+
answer_text_1.value,
|
724 |
+
is_proofreading_2.value,
|
725 |
+
answer_text_2.value,
|
726 |
+
score.value,
|
727 |
+
helphulness.value,
|
728 |
+
correcthness.value,
|
729 |
+
coherenece.value,
|
730 |
+
complexity.value,
|
731 |
+
verbosity.value,
|
732 |
+
humor.value,
|
733 |
+
creativity.value,
|
734 |
+
appropriate.value,
|
735 |
+
following_instructions.value,
|
736 |
+
politeness.value,
|
737 |
+
harmfulness.value,
|
738 |
+
hate.value,
|
739 |
+
sexual.value,
|
740 |
+
violence.value,
|
741 |
+
suicide.value,
|
742 |
+
threat.value,
|
743 |
+
gun.value,
|
744 |
+
controlled_substance.value,
|
745 |
+
climinal_planing.value,
|
746 |
+
privacy.value,
|
747 |
+
harassment.value,
|
748 |
+
profanity.value,
|
749 |
+
political_content.value,
|
750 |
+
moral_judgement.value,
|
751 |
+
typos.value,
|
752 |
+
)
|
753 |
+
|
754 |
+
# Nextデータ初期化
|
755 |
+
initialize_next_data(df)
|
756 |
+
|
757 |
+
# 5段階評価をリセット
|
758 |
+
score_reset_display()
|
759 |
+
|
760 |
+
# UI更新のための返り値
|
761 |
+
return gr.update(value=df.iloc[select_idx.value]["question"]), \
|
762 |
+
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
763 |
+
gr.update(value=df.iloc[select_idx.value]["question"]), \
|
764 |
+
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
765 |
+
gr.update(value=df.iloc[select_idx.value]["question"]), \
|
766 |
+
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
767 |
+
*score_reset_display() # スコアリセット用の返り値
|
768 |
+
|
769 |
+
gr_submit_score.click(
|
770 |
+
eval_submit,
|
771 |
+
inputs=[
|
772 |
+
gr_answer_text_3_1,
|
773 |
+
gr_score_detail,
|
774 |
+
gr_helphulness,
|
775 |
+
gr_correcthness,
|
776 |
+
gr_coherence,
|
777 |
+
gr_complexity,
|
778 |
+
gr_verbosity,
|
779 |
+
gr_humor,
|
780 |
+
gr_creativity,
|
781 |
+
gr_appropriate,
|
782 |
+
gr_following_instructions,
|
783 |
+
gr_politeness,
|
784 |
+
gr_harmfulness,
|
785 |
+
gr_hate,
|
786 |
+
gr_sexual,
|
787 |
+
gr_violence,
|
788 |
+
gr_suicide,
|
789 |
+
gr_threat,
|
790 |
+
gr_gun,
|
791 |
+
gr_controlled_substance,
|
792 |
+
gr_climinal_planing,
|
793 |
+
gr_privacy,
|
794 |
+
gr_harassment,
|
795 |
+
gr_profanity,
|
796 |
+
gr_political_content,
|
797 |
+
gr_moral_judgement,
|
798 |
+
gr_typos
|
799 |
+
],
|
800 |
+
# UI更新
|
801 |
+
outputs=[
|
802 |
+
gr_question_text_1_1,
|
803 |
+
gr_answer_text_1_1,
|
804 |
+
gr_question_text_2_1,
|
805 |
+
gr_answer_text_2_1,
|
806 |
+
gr_question_text_3_1,
|
807 |
+
gr_answer_text_3_1,
|
808 |
+
gr_score_detail,
|
809 |
+
gr_helphulness,
|
810 |
+
gr_correcthness,
|
811 |
+
gr_coherence,
|
812 |
+
gr_complexity,
|
813 |
+
gr_verbosity,
|
814 |
+
gr_humor,
|
815 |
+
gr_creativity,
|
816 |
+
gr_appropriate,
|
817 |
+
gr_following_instructions,
|
818 |
+
gr_politeness,
|
819 |
+
gr_harmfulness,
|
820 |
+
gr_hate,
|
821 |
+
gr_sexual,
|
822 |
+
gr_violence,
|
823 |
+
gr_suicide,
|
824 |
+
gr_threat,
|
825 |
+
gr_gun,
|
826 |
+
gr_controlled_substance,
|
827 |
+
gr_climinal_planing,
|
828 |
+
gr_privacy,
|
829 |
+
gr_harassment,
|
830 |
+
gr_profanity,
|
831 |
+
gr_political_content,
|
832 |
+
gr_moral_judgement,
|
833 |
+
gr_typos
|
834 |
+
],
|
835 |
+
)
|
836 |
|
837 |
# 5段階評価ボタンのクリックイベントを定義
|
838 |
def score_button_clicked(button_value):
|
839 |
+
# 状態初期化
|
840 |
good.value = False
|
841 |
bad.value = False
|
842 |
+
unknown_quality.value = False
|
843 |
+
score.value = button_value
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
844 |
|
845 |
+
# 5段階評価クリック
|
846 |
def update_annotation(
|
847 |
input_ans_1: str = None,
|
848 |
+
input_ans_2: str = None,
|
849 |
is_good: bool = None, # good/bad を表すフラグを追加
|
850 |
is_unknown: bool = None, # 「分からない」を表すフラグを追加
|
851 |
+
is_simple: bool = None,
|
852 |
score_value: int = None # 5段階評価の値、good/badの場合はNone
|
853 |
+
) -> Tuple[
|
854 |
+
gr.update,
|
855 |
+
gr.update,
|
856 |
+
gr.update,
|
857 |
+
gr.update,
|
858 |
+
gr.update,
|
859 |
+
]:
|
860 |
|
861 |
# good/bad/unknown と score の状態を更新
|
862 |
update_evaluation_state(is_good, is_unknown, score_value)
|
863 |
|
864 |
# 変更を検知 (5段階評価の場合も処理するように変更)
|
865 |
+
update_answer_state(input_ans_1, input_ans_2)
|
866 |
+
|
867 |
# 表示更新
|
868 |
# indexを進める
|
869 |
select_idx.value += 1
|
|
|
873 |
# ループさせるか、エラー処理を行う
|
874 |
if select_idx.value >= len(df):
|
875 |
select_idx.value = 0
|
876 |
+
|
877 |
+
# シンプルモードクリックで、scoreを0に設定
|
878 |
+
if is_simple == True:
|
879 |
+
score.value = 0
|
880 |
|
881 |
# データセットに追加
|
882 |
+
# 元のindex番号(dataset_id)を指定して保存
|
883 |
save_annotation_data(df)
|
884 |
|
885 |
# Nextデータ初期化
|
886 |
initialize_next_data(df)
|
887 |
+
|
888 |
+
# 5段階評価をリセット
|
889 |
+
reset_values = score_reset_display()
|
890 |
|
891 |
return gr.update(value=df.iloc[select_idx.value]["question"]), \
|
892 |
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
893 |
gr.update(value=df.iloc[select_idx.value]["question"]), \
|
894 |
+
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
895 |
+
gr.update(value=df.iloc[select_idx.value]["question"]), \
|
896 |
+
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
897 |
+
*reset_values
|
898 |
+
|
899 |
|
900 |
+
# 評価状態リセット
|
901 |
def update_evaluation_state(is_good, is_unknown, score_value):
|
902 |
if score_value is not None: # 5段階評価の場合
|
903 |
good.value = False
|
|
|
909 |
bad.value = not is_good if not is_unknown else False
|
910 |
unknown_quality.value = is_unknown # 「分からない」の状態を設定
|
911 |
|
912 |
+
# 変更を検知して値を設定
|
913 |
+
def update_answer_state(input_ans_1, input_ans_2):
|
914 |
if input_ans_1 is not None and initial_answer_text_1.value != input_ans_1:
|
915 |
is_proofreading_1.value = True
|
916 |
answer_text_1.value = input_ans_1
|
917 |
else:
|
918 |
answer_text_1.value = ""
|
919 |
|
920 |
+
# 2ターン目用
|
921 |
+
if input_ans_2 is not None and initial_answer_text_2.value != input_ans_2:
|
922 |
is_proofreading_2.value = True
|
923 |
+
answer_text_2.value = input_ans_2
|
924 |
else:
|
925 |
answer_text_2.value = ""
|
926 |
|
927 |
+
# データ保存
|
928 |
def save_annotation_data(df):
|
929 |
save_annotation(
|
930 |
select_dropdown_dataset.value,
|
931 |
# datasetIdは元のindex番号を使用
|
932 |
df.iloc[select_idx.value]['index'],
|
933 |
who.value,
|
934 |
+
unknown_quality.value,
|
935 |
good.value,
|
936 |
bad.value,
|
|
|
|
|
937 |
is_proofreading_1.value,
|
938 |
answer_text_1.value,
|
939 |
is_proofreading_2.value,
|
940 |
+
answer_text_2.value,
|
941 |
+
score.value,
|
942 |
+
helphulness.value,
|
943 |
+
correcthness.value,
|
944 |
+
coherenece.value,
|
945 |
+
complexity.value,
|
946 |
+
verbosity.value,
|
947 |
+
humor.value,
|
948 |
+
creativity.value,
|
949 |
+
appropriate.value,
|
950 |
+
following_instructions.value,
|
951 |
+
politeness.value,
|
952 |
+
harmfulness.value,
|
953 |
+
hate.value,
|
954 |
+
sexual.value,
|
955 |
+
violence.value,
|
956 |
+
suicide.value,
|
957 |
+
threat.value,
|
958 |
+
gun.value,
|
959 |
+
controlled_substance.value,
|
960 |
+
climinal_planing.value,
|
961 |
+
privacy.value,
|
962 |
+
harassment.value,
|
963 |
+
profanity.value,
|
964 |
+
political_content.value,
|
965 |
+
moral_judgement.value,
|
966 |
+
typos.value,
|
967 |
)
|
968 |
|
969 |
+
# 次データ読込時の初期化
|
970 |
def initialize_next_data(df):
|
971 |
is_proofreading_1.value = False
|
972 |
is_proofreading_2.value = False
|
973 |
initial_answer_text_1.value = df.iloc[select_idx.value]["answer"]
|
974 |
initial_answer_text_2.value = df.iloc[select_idx.value]["answer"]
|
975 |
+
|
976 |
+
# 評価不可
|
977 |
+
def unknown_click(input_ans_1, input_ans_2):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
978 |
good.value = False
|
979 |
bad.value = False
|
980 |
+
score_reset_display()
|
981 |
+
return update_annotation(
|
982 |
+
input_ans_1=input_ans_1,
|
983 |
+
input_ans_2=input_ans_2,
|
984 |
+
is_good=False,
|
985 |
+
is_unknown=True,
|
986 |
+
is_simple=True,
|
987 |
+
)
|
988 |
+
|
989 |
+
# 評価不可クリックイベント
|
990 |
+
gr_unknown_btn.click(
|
991 |
unknown_click,
|
992 |
inputs=[
|
993 |
+
gr_answer_text_1_1,
|
994 |
+
gr_answer_text_2_1
|
995 |
],
|
996 |
+
outputs=[
|
997 |
+
gr_question_text_1_1,
|
998 |
+
gr_answer_text_1_1,
|
999 |
+
gr_question_text_2_1,
|
1000 |
+
gr_answer_text_2_1,
|
1001 |
+
gr_question_text_3_1,
|
1002 |
+
gr_answer_text_3_1,
|
1003 |
+
gr_score_detail,
|
1004 |
+
gr_helphulness,
|
1005 |
+
gr_correcthness,
|
1006 |
+
gr_coherence,
|
1007 |
+
gr_complexity,
|
1008 |
+
gr_verbosity,
|
1009 |
+
gr_humor,
|
1010 |
+
gr_creativity,
|
1011 |
+
gr_appropriate,
|
1012 |
+
gr_following_instructions,
|
1013 |
+
gr_politeness,
|
1014 |
+
gr_harmfulness,
|
1015 |
+
gr_hate,
|
1016 |
+
gr_sexual,
|
1017 |
+
gr_violence,
|
1018 |
+
gr_suicide,
|
1019 |
+
gr_threat,
|
1020 |
+
gr_gun,
|
1021 |
+
gr_controlled_substance,
|
1022 |
+
gr_climinal_planing,
|
1023 |
+
gr_privacy,
|
1024 |
+
gr_harassment,
|
1025 |
+
gr_profanity,
|
1026 |
+
gr_political_content,
|
1027 |
+
gr_moral_judgement,
|
1028 |
+
gr_typos
|
1029 |
+
]
|
1030 |
)
|
1031 |
|
1032 |
+
# 良いクリック
|
1033 |
+
def good_click(input_ans_1, input_ans_2):
|
1034 |
+
score_reset_display()
|
1035 |
+
return update_annotation(
|
1036 |
+
input_ans_1=input_ans_1,
|
1037 |
+
input_ans_2=input_ans_2,
|
1038 |
+
is_good=True,
|
1039 |
+
is_unknown=False,
|
1040 |
+
is_simple=True,
|
1041 |
+
)
|
1042 |
+
|
1043 |
+
# 良いクリック
|
1044 |
+
gr_good_btn.click(
|
1045 |
+
good_click,
|
1046 |
+
inputs=[
|
1047 |
+
gr_answer_text_1_1,
|
1048 |
+
gr_answer_text_2_1
|
1049 |
+
],
|
1050 |
+
outputs=[
|
1051 |
+
gr_question_text_1_1,
|
1052 |
+
gr_answer_text_1_1,
|
1053 |
+
gr_question_text_2_1,
|
1054 |
+
gr_answer_text_2_1,
|
1055 |
+
gr_question_text_3_1,
|
1056 |
+
gr_answer_text_3_1,
|
1057 |
+
gr_score_detail,
|
1058 |
+
gr_helphulness,
|
1059 |
+
gr_correcthness,
|
1060 |
+
gr_coherence,
|
1061 |
+
gr_complexity,
|
1062 |
+
gr_verbosity,
|
1063 |
+
gr_humor,
|
1064 |
+
gr_creativity,
|
1065 |
+
gr_appropriate,
|
1066 |
+
gr_following_instructions,
|
1067 |
+
gr_politeness,
|
1068 |
+
gr_harmfulness,
|
1069 |
+
gr_hate,
|
1070 |
+
gr_sexual,
|
1071 |
+
gr_violence,
|
1072 |
+
gr_suicide,
|
1073 |
+
gr_threat,
|
1074 |
+
gr_gun,
|
1075 |
+
gr_controlled_substance,
|
1076 |
+
gr_climinal_planing,
|
1077 |
+
gr_privacy,
|
1078 |
+
gr_harassment,
|
1079 |
+
gr_profanity,
|
1080 |
+
gr_political_content,
|
1081 |
+
gr_moral_judgement,
|
1082 |
+
gr_typos
|
1083 |
+
]
|
1084 |
+
)
|
1085 |
+
|
1086 |
+
# 低評価クリック
|
1087 |
+
def bad_click(input_ans_1, input_ans_2):
|
1088 |
+
score_reset_display()
|
1089 |
+
return update_annotation(
|
1090 |
+
input_ans_1=input_ans_1,
|
1091 |
+
input_ans_2=input_ans_2,
|
1092 |
+
is_good=False,
|
1093 |
+
is_unknown=False,
|
1094 |
+
is_simple=True,
|
1095 |
+
)
|
1096 |
+
|
1097 |
+
# 低評価クリックイベント
|
1098 |
+
gr_bad_btn.click(
|
1099 |
bad_click,
|
1100 |
inputs=[
|
1101 |
+
gr_answer_text_1_1,
|
1102 |
+
gr_answer_text_2_1
|
1103 |
],
|
1104 |
+
outputs=[
|
1105 |
+
gr_question_text_1_1,
|
1106 |
+
gr_answer_text_1_1,
|
1107 |
+
gr_question_text_2_1,
|
1108 |
+
gr_answer_text_2_1,
|
1109 |
+
gr_question_text_3_1,
|
1110 |
+
gr_answer_text_3_1,
|
1111 |
+
gr_score_detail,
|
1112 |
+
gr_helphulness,
|
1113 |
+
gr_correcthness,
|
1114 |
+
gr_coherence,
|
1115 |
+
gr_complexity,
|
1116 |
+
gr_verbosity,
|
1117 |
+
gr_humor,
|
1118 |
+
gr_creativity,
|
1119 |
+
gr_appropriate,
|
1120 |
+
gr_following_instructions,
|
1121 |
+
gr_politeness,
|
1122 |
+
gr_harmfulness,
|
1123 |
+
gr_hate,
|
1124 |
+
gr_sexual,
|
1125 |
+
gr_violence,
|
1126 |
+
gr_suicide,
|
1127 |
+
gr_threat,
|
1128 |
+
gr_gun,
|
1129 |
+
gr_controlled_substance,
|
1130 |
+
gr_climinal_planing,
|
1131 |
+
gr_privacy,
|
1132 |
+
gr_harassment,
|
1133 |
+
gr_profanity,
|
1134 |
+
gr_political_content,
|
1135 |
+
gr_moral_judgement,
|
1136 |
+
gr_typos
|
1137 |
+
]
|
1138 |
)
|
1139 |
|
1140 |
# 5段階評価ボタンのクリックイベント
|
1141 |
+
gr_score_1_btn.click(lambda x: update_annotation(input_ans_1=x, is_unknown=False, score_value=1),
|
1142 |
+
inputs=[gr_answer_text_2_1],
|
1143 |
+
outputs=[
|
1144 |
+
gr_question_text_1_1,
|
1145 |
+
gr_answer_text_1_1,
|
1146 |
+
gr_question_text_2_1,
|
1147 |
+
gr_answer_text_2_1,
|
1148 |
+
gr_question_text_3_1,
|
1149 |
+
gr_answer_text_3_1,
|
1150 |
+
gr_score_detail,
|
1151 |
+
gr_helphulness,
|
1152 |
+
gr_correcthness,
|
1153 |
+
gr_coherence,
|
1154 |
+
gr_complexity,
|
1155 |
+
gr_verbosity,
|
1156 |
+
gr_humor,
|
1157 |
+
gr_creativity,
|
1158 |
+
gr_appropriate,
|
1159 |
+
gr_following_instructions,
|
1160 |
+
gr_politeness,
|
1161 |
+
gr_harmfulness,
|
1162 |
+
gr_hate,
|
1163 |
+
gr_sexual,
|
1164 |
+
gr_violence,
|
1165 |
+
gr_suicide,
|
1166 |
+
gr_threat,
|
1167 |
+
gr_gun,
|
1168 |
+
gr_controlled_substance,
|
1169 |
+
gr_climinal_planing,
|
1170 |
+
gr_privacy,
|
1171 |
+
gr_harassment,
|
1172 |
+
gr_profanity,
|
1173 |
+
gr_political_content,
|
1174 |
+
gr_moral_judgement,
|
1175 |
+
gr_typos
|
1176 |
+
])
|
1177 |
+
gr_score_2_btn.click(lambda x: update_annotation(input_ans_1=x, is_unknown=False, score_value=2),
|
1178 |
+
inputs=[gr_answer_text_2_1],
|
1179 |
+
outputs=[
|
1180 |
+
gr_question_text_1_1,
|
1181 |
+
gr_answer_text_1_1,
|
1182 |
+
gr_question_text_2_1,
|
1183 |
+
gr_answer_text_2_1,
|
1184 |
+
gr_question_text_3_1,
|
1185 |
+
gr_answer_text_3_1,
|
1186 |
+
gr_score_detail,
|
1187 |
+
gr_helphulness,
|
1188 |
+
gr_correcthness,
|
1189 |
+
gr_coherence,
|
1190 |
+
gr_complexity,
|
1191 |
+
gr_verbosity,
|
1192 |
+
gr_humor,
|
1193 |
+
gr_creativity,
|
1194 |
+
gr_appropriate,
|
1195 |
+
gr_following_instructions,
|
1196 |
+
gr_politeness,
|
1197 |
+
gr_harmfulness,
|
1198 |
+
gr_hate,
|
1199 |
+
gr_sexual,
|
1200 |
+
gr_violence,
|
1201 |
+
gr_suicide,
|
1202 |
+
gr_threat,
|
1203 |
+
gr_gun,
|
1204 |
+
gr_controlled_substance,
|
1205 |
+
gr_climinal_planing,
|
1206 |
+
gr_privacy,
|
1207 |
+
gr_harassment,
|
1208 |
+
gr_profanity,
|
1209 |
+
gr_political_content,
|
1210 |
+
gr_moral_judgement,
|
1211 |
+
gr_typos
|
1212 |
+
])
|
1213 |
+
gr_score_3_btn.click(lambda x: update_annotation(input_ans_1=x, is_unknown=False, score_value=3),
|
1214 |
+
inputs=[gr_answer_text_2_1],
|
1215 |
+
outputs=[
|
1216 |
+
gr_question_text_1_1,
|
1217 |
+
gr_answer_text_1_1,
|
1218 |
+
gr_question_text_2_1,
|
1219 |
+
gr_answer_text_2_1,
|
1220 |
+
gr_question_text_3_1,
|
1221 |
+
gr_answer_text_3_1,
|
1222 |
+
gr_score_detail,
|
1223 |
+
gr_helphulness,
|
1224 |
+
gr_correcthness,
|
1225 |
+
gr_coherence,
|
1226 |
+
gr_complexity,
|
1227 |
+
gr_verbosity,
|
1228 |
+
gr_humor,
|
1229 |
+
gr_creativity,
|
1230 |
+
gr_appropriate,
|
1231 |
+
gr_following_instructions,
|
1232 |
+
gr_politeness,
|
1233 |
+
gr_harmfulness,
|
1234 |
+
gr_hate,
|
1235 |
+
gr_sexual,
|
1236 |
+
gr_violence,
|
1237 |
+
gr_suicide,
|
1238 |
+
gr_threat,
|
1239 |
+
gr_gun,
|
1240 |
+
gr_controlled_substance,
|
1241 |
+
gr_climinal_planing,
|
1242 |
+
gr_privacy,
|
1243 |
+
gr_harassment,
|
1244 |
+
gr_profanity,
|
1245 |
+
gr_political_content,
|
1246 |
+
gr_moral_judgement,
|
1247 |
+
gr_typos
|
1248 |
+
])
|
1249 |
+
gr_score_4_btn.click(lambda x: update_annotation(input_ans_1=x, is_unknown=False, score_value=4),
|
1250 |
+
inputs=[gr_answer_text_2_1],
|
1251 |
+
outputs=[
|
1252 |
+
gr_question_text_1_1,
|
1253 |
+
gr_answer_text_1_1,
|
1254 |
+
gr_question_text_2_1,
|
1255 |
+
gr_answer_text_2_1,
|
1256 |
+
gr_question_text_3_1,
|
1257 |
+
gr_answer_text_3_1,
|
1258 |
+
gr_score_detail,
|
1259 |
+
gr_helphulness,
|
1260 |
+
gr_correcthness,
|
1261 |
+
gr_coherence,
|
1262 |
+
gr_complexity,
|
1263 |
+
gr_verbosity,
|
1264 |
+
gr_humor,
|
1265 |
+
gr_creativity,
|
1266 |
+
gr_appropriate,
|
1267 |
+
gr_following_instructions,
|
1268 |
+
gr_politeness,
|
1269 |
+
gr_harmfulness,
|
1270 |
+
gr_hate,
|
1271 |
+
gr_sexual,
|
1272 |
+
gr_violence,
|
1273 |
+
gr_suicide,
|
1274 |
+
gr_threat,
|
1275 |
+
gr_gun,
|
1276 |
+
gr_controlled_substance,
|
1277 |
+
gr_climinal_planing,
|
1278 |
+
gr_privacy,
|
1279 |
+
gr_harassment,
|
1280 |
+
gr_profanity,
|
1281 |
+
gr_political_content,
|
1282 |
+
gr_moral_judgement,
|
1283 |
+
gr_typos
|
1284 |
+
])
|
1285 |
+
gr_score_5_btn.click(lambda x: update_annotation(input_ans_1=x, is_unknown=False, score_value=5),
|
1286 |
+
inputs=[gr_answer_text_2_1],
|
1287 |
+
outputs=[
|
1288 |
+
gr_question_text_1_1,
|
1289 |
+
gr_answer_text_1_1,
|
1290 |
+
gr_question_text_2_1,
|
1291 |
+
gr_answer_text_2_1,
|
1292 |
+
gr_question_text_3_1,
|
1293 |
+
gr_answer_text_3_1,
|
1294 |
+
gr_score_detail,
|
1295 |
+
gr_helphulness,
|
1296 |
+
gr_correcthness,
|
1297 |
+
gr_coherence,
|
1298 |
+
gr_complexity,
|
1299 |
+
gr_verbosity,
|
1300 |
+
gr_humor,
|
1301 |
+
gr_creativity,
|
1302 |
+
gr_appropriate,
|
1303 |
+
gr_following_instructions,
|
1304 |
+
gr_politeness,
|
1305 |
+
gr_harmfulness,
|
1306 |
+
gr_hate,
|
1307 |
+
gr_sexual,
|
1308 |
+
gr_violence,
|
1309 |
+
gr_suicide,
|
1310 |
+
gr_threat,
|
1311 |
+
gr_gun,
|
1312 |
+
gr_controlled_substance,
|
1313 |
+
gr_climinal_planing,
|
1314 |
+
gr_privacy,
|
1315 |
+
gr_harassment,
|
1316 |
+
gr_profanity,
|
1317 |
+
gr_political_content,
|
1318 |
+
gr_moral_judgement,
|
1319 |
+
gr_typos
|
1320 |
+
])
|
1321 |
+
|
1322 |
+
# Scoreリセット
|
1323 |
+
def score_reset_display():
|
1324 |
+
# gr.State
|
1325 |
+
score.value = 3
|
1326 |
+
helphulness.value = 0
|
1327 |
+
correcthness.value = 0
|
1328 |
+
coherenece.value = 0
|
1329 |
+
complexity.value = 0
|
1330 |
+
verbosity.value = 0
|
1331 |
+
humor.value = 0
|
1332 |
+
creativity.value = 0
|
1333 |
+
appropriate.value = 0
|
1334 |
+
following_instructions.value = 0
|
1335 |
+
politeness.value = 0
|
1336 |
+
harmfulness.value = 0
|
1337 |
+
hate.value = 0
|
1338 |
+
sexual.value = 0
|
1339 |
+
violence.value = 0
|
1340 |
+
suicide.value = 0
|
1341 |
+
threat.value = 0
|
1342 |
+
gun.value = 0
|
1343 |
+
controlled_substance.vlaue = 0
|
1344 |
+
climinal_planing.value = 0
|
1345 |
+
privacy.value = 0
|
1346 |
+
harassment.value = 0
|
1347 |
+
profanity.value = 0
|
1348 |
+
political_content.value = 0
|
1349 |
+
moral_judgement.value = 0
|
1350 |
+
typos.value = 0
|
1351 |
+
|
1352 |
+
return 3,0,0,0,0,0,0,0,0,0,0,0, \
|
1353 |
+
"不明", "不明", "不明", "不明", "不明", \
|
1354 |
+
"不明", "不明", "不明", "不明", "不明", \
|
1355 |
+
"不明", "不明", "不明", "不明"
|
1356 |
+
|
1357 |
+
# Scoreリセット
|
1358 |
+
gr_score_reset.click(
|
1359 |
+
score_reset_display,
|
1360 |
+
inputs=[],
|
1361 |
+
outputs=[
|
1362 |
+
gr_score_detail,
|
1363 |
+
gr_helphulness,
|
1364 |
+
gr_correcthness,
|
1365 |
+
gr_coherence,
|
1366 |
+
gr_complexity,
|
1367 |
+
gr_verbosity,
|
1368 |
+
gr_humor,
|
1369 |
+
gr_creativity,
|
1370 |
+
gr_appropriate,
|
1371 |
+
gr_following_instructions,
|
1372 |
+
gr_politeness,
|
1373 |
+
gr_harmfulness,
|
1374 |
+
gr_hate,
|
1375 |
+
gr_sexual,
|
1376 |
+
gr_violence,
|
1377 |
+
gr_suicide,
|
1378 |
+
gr_threat,
|
1379 |
+
gr_gun,
|
1380 |
+
gr_controlled_substance,
|
1381 |
+
gr_climinal_planing,
|
1382 |
+
gr_privacy,
|
1383 |
+
gr_harassment,
|
1384 |
+
gr_profanity,
|
1385 |
+
gr_political_content,
|
1386 |
+
gr_moral_judgement,
|
1387 |
+
gr_typos
|
1388 |
+
]
|
1389 |
+
)
|
1390 |
+
|
1391 |
+
# データ読込
|
1392 |
+
gr_data_load_btn.click(
|
1393 |
+
dataset_load_fn,
|
1394 |
+
inputs=None,
|
1395 |
+
# textをセットし、intractiveをTrueにする (UI更新)
|
1396 |
+
outputs=[
|
1397 |
+
gr_question_text_1_1,
|
1398 |
+
gr_answer_text_1_1,
|
1399 |
+
gr_question_text_2_1,
|
1400 |
+
gr_answer_text_2_1,
|
1401 |
+
gr_question_text_3_1,
|
1402 |
+
gr_answer_text_3_1,
|
1403 |
+
gr_answer_text_1_1,
|
1404 |
+
gr_answer_text_2_1,
|
1405 |
+
gr_answer_text_3_1,
|
1406 |
+
gr_unknown_btn,
|
1407 |
+
gr_good_btn,
|
1408 |
+
gr_bad_btn,
|
1409 |
+
gr_score_5_btn,
|
1410 |
+
gr_score_4_btn,
|
1411 |
+
gr_score_3_btn,
|
1412 |
+
gr_score_2_btn,
|
1413 |
+
gr_score_1_btn,
|
1414 |
+
gr_submit_score,
|
1415 |
+
gr_score_detail,
|
1416 |
+
gr_score_reset,
|
1417 |
+
gr_helphulness,
|
1418 |
+
gr_correcthness,
|
1419 |
+
gr_coherence,
|
1420 |
+
gr_complexity,
|
1421 |
+
gr_verbosity,
|
1422 |
+
gr_humor,
|
1423 |
+
gr_creativity,
|
1424 |
+
gr_appropriate,
|
1425 |
+
gr_following_instructions,
|
1426 |
+
gr_politeness,
|
1427 |
+
gr_harmfulness,
|
1428 |
+
gr_hate,
|
1429 |
+
gr_sexual,
|
1430 |
+
gr_violence,
|
1431 |
+
gr_suicide,
|
1432 |
+
gr_threat,
|
1433 |
+
gr_gun,
|
1434 |
+
gr_controlled_substance,
|
1435 |
+
gr_climinal_planing,
|
1436 |
+
gr_privacy,
|
1437 |
+
gr_harassment,
|
1438 |
+
gr_profanity,
|
1439 |
+
gr_political_content,
|
1440 |
+
gr_moral_judgement,
|
1441 |
+
gr_typos
|
1442 |
+
]
|
1443 |
+
)
|
1444 |
+
|
1445 |
# TODO Tab切り替えで、アノテ済みの一覧を表示する
|
1446 |
# with gr.Tab("アノテ済みデータセット(管理画面)"):
|
1447 |
# タブを切り替えた時にデータ表示を更新する
|
run.py
CHANGED
@@ -28,14 +28,14 @@ else:
|
|
28 |
print("Warning: HF_TOKEN not found. Please set it in your Space secrets.")
|
29 |
|
30 |
# HFデータセット アップロード先
|
31 |
-
# (
|
32 |
-
|
33 |
"team-hatakeyama-phase2/annotation_tanuki_phase2",
|
34 |
"kevineen/Tanuki-Phase2-annotation-dataset", # 出力テスト
|
35 |
]
|
36 |
|
37 |
# アノテーション対象データセット
|
38 |
-
|
39 |
"hatakeyama-llm-team/AutoGeneratedJapaneseQA",
|
40 |
"hatakeyama-llm-team/AutoGeneratedJapaneseQA-other",
|
41 |
"kanhatakeyama/ChatbotArenaJaMixtral8x22b",
|
@@ -46,9 +46,11 @@ annotation_dataset_list = [
|
|
46 |
# "hatakeyama-llm-team/WikiBookJa",
|
47 |
# "kanhatakeyama/AutoWikiQA",
|
48 |
# "susumuota/SyntheticTextWikiTranslate-askllm-v1", # Ask-LLM 翻訳
|
|
|
|
|
49 |
]
|
50 |
|
51 |
-
|
52 |
# マルチターン 未対応
|
53 |
"kanhatakeyama/AutoMultiTurnByMixtral8x22b",
|
54 |
]
|
@@ -60,56 +62,152 @@ is_selected_dataset = gr.State(False)
|
|
60 |
is_loaded_dataset = gr.State(False)
|
61 |
|
62 |
# 選択中のデータセットリスト
|
63 |
-
dropdown_dataset_list = gr.State(value=
|
|
|
64 |
# 現在の対象データセット 初期値は"hatakeyama-llm-team/AutoGeneratedJapaneseQA",
|
65 |
select_dropdown_dataset = gr.State(dropdown_dataset_list.value[0])
|
66 |
select_dataset = gr.State(None) # 現在のデータセット
|
67 |
select_dataset_total_len = gr.State(0) # 現在のデータセットの長さ
|
68 |
select_idx = gr.State(0) # 現在のインデックス (ランダムモードにするなら不要?
|
69 |
-
random_mode = gr.State(False)
|
70 |
|
71 |
# 回答者がアノテーションしたデータセット
|
72 |
annotated_dataset = gr.State(
|
73 |
pd.DataFrame({
|
74 |
-
'dataset_name': [],
|
75 |
-
'dataset_id': [],
|
76 |
-
'who': [],
|
77 |
-
'
|
78 |
-
'
|
79 |
-
'
|
80 |
-
'
|
81 |
-
'is_proofreading_1': [],
|
82 |
'answer_text_1': [],
|
83 |
-
'is_proofreading_2': [], #
|
84 |
-
'answer_text_2': [], #
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
})
|
86 |
)
|
87 |
|
88 |
-
|
89 |
-
initial_answer_text_2 = gr.State("") # 回答2を整形したかチェック用
|
90 |
-
|
91 |
-
is_dataset_loaded = gr.State(False)
|
92 |
|
93 |
you_dataset_id = gr.State(0) # 回答者がアノテーションしているデータのID
|
94 |
dataset_name = gr.State("") # 編集に使用したデータセット名
|
95 |
dataset_id = gr.State(0) # 加工元データセットのindex
|
96 |
who = gr.State("") # アノテーション者名
|
|
|
|
|
|
|
97 |
good = gr.State(False) # 良
|
98 |
bad = gr.State(False) # 悪
|
99 |
-
|
100 |
-
|
|
|
|
|
101 |
is_proofreading_1 = gr.State(False) # 回答1を整形したか_1
|
102 |
answer_text_1 = gr.State("") # answer_1 回答
|
103 |
is_proofreading_2 = gr.State(False) # 回答2を整形したか_2
|
104 |
answer_text_2 = gr.State("") # answer_2 回答
|
105 |
|
106 |
-
#
|
107 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
def dataset_load_fn() -> Tuple[
|
109 |
str,
|
110 |
str,
|
111 |
str,
|
112 |
str,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
113 |
gr.update,
|
114 |
gr.update,
|
115 |
gr.update,
|
@@ -118,9 +216,11 @@ def dataset_load_fn() -> Tuple[
|
|
118 |
gr.update,
|
119 |
gr.update,
|
120 |
gr.update,
|
121 |
-
gr.update
|
|
|
|
|
122 |
|
123 |
-
is_dataset_loaded.value = False #
|
124 |
|
125 |
select_dataset.value = load_dataset(
|
126 |
select_dropdown_dataset.value
|
@@ -129,6 +229,7 @@ def dataset_load_fn() -> Tuple[
|
|
129 |
# DatasetオブジェクトをPandas DataFrameに変換
|
130 |
df = select_dataset.value["train"].to_pandas()
|
131 |
|
|
|
132 |
# index列を追加し、シャッフル
|
133 |
df = df.reset_index(drop=False) # 元のindexを保持
|
134 |
df = df.sample(frac=1).reset_index(drop=True) # シャッフル
|
@@ -138,7 +239,7 @@ def dataset_load_fn() -> Tuple[
|
|
138 |
select_dataset_total_len.value = len(df) # 長さを取得
|
139 |
is_dataset_loaded.value = True # ロード完了
|
140 |
|
141 |
-
#
|
142 |
initial_answer_text_1.value = df.iloc[select_idx.value]["answer"]
|
143 |
initial_answer_text_2.value = df.iloc[select_idx.value]["answer"]
|
144 |
|
@@ -146,6 +247,39 @@ def dataset_load_fn() -> Tuple[
|
|
146 |
df.iloc[select_idx.value]["answer"], \
|
147 |
df.iloc[select_idx.value]["question"], \
|
148 |
df.iloc[select_idx.value]["answer"], \
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
149 |
gr.update(interactive=True), \
|
150 |
gr.update(interactive=True), \
|
151 |
gr.update(interactive=True), \
|
@@ -164,7 +298,7 @@ annotation_file = Path("user_annotation/") / f"data_{uuid.uuid4()}.json"
|
|
164 |
annotated_folder = annotation_file.parent
|
165 |
|
166 |
scheduler = CommitScheduler(
|
167 |
-
repo_id=
|
168 |
repo_type="dataset",
|
169 |
folder_path=annotated_folder,
|
170 |
path_in_repo="data", # Spaceの場合の保存先フォルダー
|
@@ -172,19 +306,58 @@ scheduler = CommitScheduler(
|
|
172 |
every=5, # 5分毎にアップロード HuggingFAce_Documentの最低推奨値
|
173 |
)
|
174 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
175 |
# CommitScheduler (HFへのデータアップロード
|
176 |
def save_annotation(
|
177 |
dataset_name: str,
|
178 |
dataset_id: int,
|
179 |
who: str,
|
|
|
180 |
good: bool,
|
181 |
bad: bool,
|
182 |
-
unknown_quality: bool,
|
183 |
-
score: int,
|
184 |
is_proofreading_1: bool,
|
185 |
answer_text_1: str,
|
186 |
is_proofreading_2: bool,
|
187 |
-
answer_text_2: str
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
188 |
|
189 |
annotated_dataset.value = pd.concat([
|
190 |
annotated_dataset.value,
|
@@ -192,33 +365,82 @@ def save_annotation(
|
|
192 |
'dataset_name': [dataset_name],
|
193 |
'dataset_id': [dataset_id],
|
194 |
'who': [who],
|
|
|
195 |
'good': [good],
|
196 |
'bad': [bad],
|
197 |
-
'unknown_quality': [unknown_quality],
|
198 |
-
'score': [score],
|
199 |
'is_proofreading_1': [is_proofreading_1],
|
200 |
"answer_text_1": [answer_text_1],
|
201 |
'is_proofreading_2': [is_proofreading_2],
|
202 |
-
'answer_text_2': [answer_text_2]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
203 |
})], ignore_index=True).reset_index(drop=True)
|
204 |
|
205 |
# 書き込み
|
206 |
with scheduler.lock:
|
207 |
with annotation_file.open("a", encoding='utf-8') as f:
|
208 |
data_to_write = {
|
209 |
-
# "id": , CommitScheduler
|
210 |
"datetime": str(datetime.datetime.now().isoformat()),
|
211 |
"dataset_name": dataset_name,
|
212 |
"dataset_id": int(dataset_id),
|
213 |
"who": who,
|
|
|
214 |
"good": good,
|
215 |
"bad": bad,
|
216 |
-
"unknown_quality": unknown_quality, # 新しいデータフィールドを追加
|
217 |
-
"score": score,
|
218 |
"is_proofreading_1": is_proofreading_1,
|
219 |
"answer_text_1": answer_text_1,
|
220 |
"is_proofreading_2": is_proofreading_2,
|
221 |
"answer_text_2": answer_text_2,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
222 |
}
|
223 |
f.write(json.dumps(data_to_write, ensure_ascii=False))
|
224 |
f.write("\n")
|
@@ -230,7 +452,6 @@ def save_annotation(
|
|
230 |
|
231 |
# ユーザー名表示
|
232 |
|
233 |
-
|
234 |
def hello(profile: gr.OAuthProfile | None) -> Tuple[str, str]:
|
235 |
if profile is None:
|
236 |
return "プライベートデータセット取得のためにログインしてください。", who.value
|
@@ -242,30 +463,40 @@ def hello(profile: gr.OAuthProfile | None) -> Tuple[str, str]:
|
|
242 |
theme_ = gr.themes.Default()
|
243 |
|
244 |
# 後のCSSデザイン変更用
|
245 |
-
|
246 |
-
|
247 |
def load_css():
|
248 |
with open("style.css", "r") as file:
|
249 |
css_content = file.read()
|
250 |
return css_content
|
251 |
|
252 |
-
|
253 |
# Gradio 画面 ============================================
|
254 |
with gr.Blocks(theme=theme_, css=load_css()) as demo:
|
255 |
|
256 |
gr.Markdown("# データセット アノテーション for Tanuki (Phase2)")
|
257 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
258 |
|
259 |
with gr.Tab("アノテーション (シングルターン)"):
|
260 |
|
261 |
-
|
262 |
-
|
263 |
-
gr.LoginButton(value="HuggingFace ログイン",
|
264 |
-
|
265 |
-
|
266 |
# ユーザー名
|
267 |
-
gr_profile_name = gr.Markdown()
|
268 |
-
demo.load(hello, inputs=None, outputs=[gr_profile_name, who])
|
269 |
|
270 |
with gr.Row():
|
271 |
|
@@ -288,76 +519,353 @@ with gr.Blocks(theme=theme_, css=load_css()) as demo:
|
|
288 |
gr_data_load_btn = gr.Button("② データセットを読み込む")
|
289 |
|
290 |
with gr.Column() as content_column:
|
|
|
291 |
with gr.Tab("③ シンプル(良・悪)"):
|
292 |
with gr.Column():
|
293 |
with gr.Row(equal_height=True):
|
294 |
-
|
295 |
-
|
296 |
"分からない", interactive=False) # 「分からない」ボタンを追加
|
297 |
-
|
298 |
|
299 |
-
|
300 |
label="質問: ", lines=5, interactive=False)
|
301 |
|
302 |
-
|
303 |
-
label="回答:
|
304 |
lines=20,
|
305 |
-
interactive=
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
306 |
|
307 |
-
|
308 |
|
309 |
-
|
310 |
-
|
|
|
311 |
|
312 |
-
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
318 |
|
319 |
-
|
320 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
321 |
|
322 |
# 5段階評価ボタンのクリックイベントを定義
|
323 |
def score_button_clicked(button_value):
|
|
|
324 |
good.value = False
|
325 |
bad.value = False
|
326 |
-
unknown_quality.value = False
|
327 |
-
score.value = button_value
|
328 |
-
|
329 |
-
gr_data_load_btn.click(
|
330 |
-
dataset_load_fn,
|
331 |
-
inputs=None,
|
332 |
-
outputs=[gr_question_text_1,
|
333 |
-
gr_answer_text_1,
|
334 |
-
gr_question_text_3,
|
335 |
-
gr_answer_text_3,
|
336 |
-
good_btn,
|
337 |
-
unknown_btn, # 「分からない」ボタンを追加
|
338 |
-
bad_btn,
|
339 |
-
gr_score_1,
|
340 |
-
gr_score_2,
|
341 |
-
gr_score_3,
|
342 |
-
gr_score_4,
|
343 |
-
gr_score_5,
|
344 |
-
]
|
345 |
-
)
|
346 |
|
|
|
347 |
def update_annotation(
|
348 |
input_ans_1: str = None,
|
349 |
-
|
350 |
is_good: bool = None, # good/bad を表すフラグを追加
|
351 |
is_unknown: bool = None, # 「分からない」を表すフラグを追加
|
|
|
352 |
score_value: int = None # 5段階評価の値、good/badの場合はNone
|
353 |
-
) -> Tuple[
|
|
|
|
|
|
|
|
|
|
|
|
|
354 |
|
355 |
# good/bad/unknown と score の状態を更新
|
356 |
update_evaluation_state(is_good, is_unknown, score_value)
|
357 |
|
358 |
# 変更を検知 (5段階評価の場合も処理するように変更)
|
359 |
-
update_answer_state(input_ans_1,
|
360 |
-
|
361 |
# 表示更新
|
362 |
# indexを進める
|
363 |
select_idx.value += 1
|
@@ -367,19 +875,31 @@ with gr.Blocks(theme=theme_, css=load_css()) as demo:
|
|
367 |
# ループさせるか、エラー処理を行う
|
368 |
if select_idx.value >= len(df):
|
369 |
select_idx.value = 0
|
|
|
|
|
|
|
|
|
370 |
|
371 |
# データセットに追加
|
372 |
-
# 元のindex番号(dataset_id)を指定して保存
|
373 |
save_annotation_data(df)
|
374 |
|
375 |
# Nextデータ初期化
|
376 |
initialize_next_data(df)
|
|
|
|
|
|
|
377 |
|
378 |
return gr.update(value=df.iloc[select_idx.value]["question"]), \
|
379 |
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
380 |
gr.update(value=df.iloc[select_idx.value]["question"]), \
|
381 |
-
gr.update(value=df.iloc[select_idx.value]["answer"])
|
|
|
|
|
|
|
|
|
382 |
|
|
|
383 |
def update_evaluation_state(is_good, is_unknown, score_value):
|
384 |
if score_value is not None: # 5段階評価の場合
|
385 |
good.value = False
|
@@ -391,100 +911,539 @@ with gr.Blocks(theme=theme_, css=load_css()) as demo:
|
|
391 |
bad.value = not is_good if not is_unknown else False
|
392 |
unknown_quality.value = is_unknown # 「分からない」の状態を設定
|
393 |
|
394 |
-
|
|
|
395 |
if input_ans_1 is not None and initial_answer_text_1.value != input_ans_1:
|
396 |
is_proofreading_1.value = True
|
397 |
answer_text_1.value = input_ans_1
|
398 |
else:
|
399 |
answer_text_1.value = ""
|
400 |
|
401 |
-
|
|
|
402 |
is_proofreading_2.value = True
|
403 |
-
answer_text_2.value =
|
404 |
else:
|
405 |
answer_text_2.value = ""
|
406 |
|
|
|
407 |
def save_annotation_data(df):
|
408 |
save_annotation(
|
409 |
select_dropdown_dataset.value,
|
410 |
# datasetIdは元のindex番号を使用
|
411 |
df.iloc[select_idx.value]['index'],
|
412 |
who.value,
|
|
|
413 |
good.value,
|
414 |
bad.value,
|
415 |
-
unknown_quality.value, # 「分からない」の状態を追加
|
416 |
-
score.value,
|
417 |
is_proofreading_1.value,
|
418 |
answer_text_1.value,
|
419 |
is_proofreading_2.value,
|
420 |
-
answer_text_2.value
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
421 |
)
|
422 |
|
|
|
423 |
def initialize_next_data(df):
|
424 |
is_proofreading_1.value = False
|
425 |
is_proofreading_2.value = False
|
426 |
initial_answer_text_1.value = df.iloc[select_idx.value]["answer"]
|
427 |
initial_answer_text_2.value = df.iloc[select_idx.value]["answer"]
|
428 |
-
|
429 |
-
|
430 |
-
|
431 |
-
|
432 |
-
good_btn.click(
|
433 |
-
good_click,
|
434 |
-
inputs=[
|
435 |
-
gr_answer_text_1,
|
436 |
-
gr_answer_text_3
|
437 |
-
],
|
438 |
-
outputs=[gr_question_text_1,
|
439 |
-
gr_answer_text_1,
|
440 |
-
gr_question_text_3,
|
441 |
-
gr_answer_text_3]
|
442 |
-
)
|
443 |
-
|
444 |
-
def unknown_click(input_ans_1, input_ans_3):
|
445 |
good.value = False
|
446 |
bad.value = False
|
447 |
-
|
448 |
-
|
449 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
450 |
unknown_click,
|
451 |
inputs=[
|
452 |
-
|
453 |
-
|
454 |
],
|
455 |
-
outputs=[
|
456 |
-
|
457 |
-
|
458 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
459 |
)
|
460 |
|
461 |
-
|
462 |
-
|
463 |
-
|
464 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
465 |
bad_click,
|
466 |
inputs=[
|
467 |
-
|
468 |
-
|
469 |
],
|
470 |
-
outputs=[
|
471 |
-
|
472 |
-
|
473 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
474 |
)
|
475 |
|
476 |
# 5段階評価ボタンのクリックイベント
|
477 |
-
|
478 |
-
|
479 |
-
|
480 |
-
|
481 |
-
|
482 |
-
|
483 |
-
|
484 |
-
|
485 |
-
|
486 |
-
|
487 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
488 |
# TODO Tab切り替えで、アノテ済みの一覧を表示する
|
489 |
# with gr.Tab("アノテ済みデータセット(管理画面)"):
|
490 |
# タブを切り替えた時にデータ表示を更新する
|
|
|
28 |
print("Warning: HF_TOKEN not found. Please set it in your Space secrets.")
|
29 |
|
30 |
# HFデータセット アップロード先
|
31 |
+
# (切り替えてテストする用に配列)
|
32 |
+
OUTPUT_DATASET = [
|
33 |
"team-hatakeyama-phase2/annotation_tanuki_phase2",
|
34 |
"kevineen/Tanuki-Phase2-annotation-dataset", # 出力テスト
|
35 |
]
|
36 |
|
37 |
# アノテーション対象データセット
|
38 |
+
ANNOTATION_DATASET = [
|
39 |
"hatakeyama-llm-team/AutoGeneratedJapaneseQA",
|
40 |
"hatakeyama-llm-team/AutoGeneratedJapaneseQA-other",
|
41 |
"kanhatakeyama/ChatbotArenaJaMixtral8x22b",
|
|
|
46 |
# "hatakeyama-llm-team/WikiBookJa",
|
47 |
# "kanhatakeyama/AutoWikiQA",
|
48 |
# "susumuota/SyntheticTextWikiTranslate-askllm-v1", # Ask-LLM 翻訳
|
49 |
+
|
50 |
+
# Nomotron データ (将来用
|
51 |
]
|
52 |
|
53 |
+
MULTITURN_ANNOTATION_DATASET = [
|
54 |
# マルチターン 未対応
|
55 |
"kanhatakeyama/AutoMultiTurnByMixtral8x22b",
|
56 |
]
|
|
|
62 |
is_loaded_dataset = gr.State(False)
|
63 |
|
64 |
# 選択中のデータセットリスト
|
65 |
+
dropdown_dataset_list = gr.State(value=ANNOTATION_DATASET)
|
66 |
+
|
67 |
# 現在の対象データセット 初期値は"hatakeyama-llm-team/AutoGeneratedJapaneseQA",
|
68 |
select_dropdown_dataset = gr.State(dropdown_dataset_list.value[0])
|
69 |
select_dataset = gr.State(None) # 現在のデータセット
|
70 |
select_dataset_total_len = gr.State(0) # 現在のデータセットの長さ
|
71 |
select_idx = gr.State(0) # 現在のインデックス (ランダムモードにするなら不要?
|
72 |
+
# random_mode = gr.State(False) # ランダムモード
|
73 |
|
74 |
# 回答者がアノテーションしたデータセット
|
75 |
annotated_dataset = gr.State(
|
76 |
pd.DataFrame({
|
77 |
+
'dataset_name': [], # 対象データセット
|
78 |
+
'dataset_id': [], # データセットindex
|
79 |
+
'who': [], # アノテーション者名
|
80 |
+
'unknown_quality': [], # 不明能
|
81 |
+
'good': [], # 良
|
82 |
+
'bad': [], # 悪
|
83 |
+
'is_proofreading_1': [], # テキストに修正があったか
|
|
|
84 |
'answer_text_1': [],
|
85 |
+
'is_proofreading_2': [], # 2ターン目用
|
86 |
+
'answer_text_2': [], # 2ターン目用
|
87 |
+
|
88 |
+
# 5段階評価
|
89 |
+
'score': [], # 総合 (5段階評価)
|
90 |
+
'helphulness': [], # 有用性 (5段階評価)
|
91 |
+
'correcthness': [], # 正確さ (5段階評価)
|
92 |
+
'coherence': [], # 一貫性 (5段階評価)
|
93 |
+
'complexity': [], # 複雑さ (5段階評価)
|
94 |
+
'verbosity': [], # 冗長性 (5段階評価)
|
95 |
+
'humor': [], # ユーモア (SteerLM)
|
96 |
+
'creativity': [], # 創造性 (SteerLM)
|
97 |
+
'appropriate': [], # 適切性 (SteerLMでは不適)
|
98 |
+
'following_instructions': [], # 指示が合った場合、従っているか (SteerLMのfails_taskに該当)
|
99 |
+
'politeness': [], # 礼儀正しさ (minnadechat)
|
100 |
+
'harmfulness': [], # 有性 (minnadechat)
|
101 |
+
|
102 |
+
# 0: 分からない(未設定) 1: 無 2: 有
|
103 |
+
'typos': [], # 誤字脱字 (minnadechat)
|
104 |
+
'hate': [], # ヘイト
|
105 |
+
'sexual': [], # 性的
|
106 |
+
'violence': [], # 暴力
|
107 |
+
'suicide': [], # 自殺
|
108 |
+
'threat': [], # 犯罪
|
109 |
+
'gun': [], # 重火器
|
110 |
+
'controlled_substance': [], # 規制対象物質
|
111 |
+
'climinal_planing': [], # 犯罪計画
|
112 |
+
'privacy': [], # 個人情報
|
113 |
+
'harassment': [], # ハラスメント
|
114 |
+
'profanity': [], # 冒涜
|
115 |
+
'political_content': [], # 政治的内容 (SteerLM)
|
116 |
+
'moral_judgement': [], # 非道徳 (SteerLM)
|
117 |
})
|
118 |
)
|
119 |
|
120 |
+
is_dataset_loaded = gr.State(False) # データセットロード状態
|
|
|
|
|
|
|
121 |
|
122 |
you_dataset_id = gr.State(0) # 回答者がアノテーションしているデータのID
|
123 |
dataset_name = gr.State("") # 編集に使用したデータセット名
|
124 |
dataset_id = gr.State(0) # 加工元データセットのindex
|
125 |
who = gr.State("") # アノテーション者名
|
126 |
+
|
127 |
+
# シンプルモード
|
128 |
+
unknown_quality = gr.State(False) # 分からない(未設定)
|
129 |
good = gr.State(False) # 良
|
130 |
bad = gr.State(False) # 悪
|
131 |
+
|
132 |
+
initial_answer_text_1 = gr.State("") # 回��1を整形したかチェック用
|
133 |
+
initial_answer_text_2 = gr.State("") # 回答2を整形したかチェック用
|
134 |
+
|
135 |
is_proofreading_1 = gr.State(False) # 回答1を整形したか_1
|
136 |
answer_text_1 = gr.State("") # answer_1 回答
|
137 |
is_proofreading_2 = gr.State(False) # 回答2を整形したか_2
|
138 |
answer_text_2 = gr.State("") # answer_2 回答
|
139 |
|
140 |
+
# (5段階評価)
|
141 |
+
# 0は判断していない、又は判断が難しいという特徴量にする
|
142 |
+
score = gr.State(3) # 総合スコア 初期値は3 (最低限必須)
|
143 |
+
helphulness = gr.State(0) # 有用性
|
144 |
+
correcthness = gr.State(0) # 正確さ
|
145 |
+
coherenece = gr.State(0) # 一貫性
|
146 |
+
complexity = gr.State(0) # 複雑さ
|
147 |
+
verbosity = gr.State(0) # 冗長性
|
148 |
+
humor = gr.State(0) # ユーモア (SteerLM)
|
149 |
+
creativity = gr.State(0) # 創造性 (SteerLM)
|
150 |
+
appropriate = gr.State(0) # 適切性 (SteerLMではnot_appropriate)
|
151 |
+
following_instructions = gr.State(0) # 指示への忠実度 (SteerLMのfails_task)
|
152 |
+
politeness = gr.State(0) # 礼儀正しさ (minnadechat)
|
153 |
+
harmfulness = gr.State(0) # 有度 (minnadechat)
|
154 |
+
|
155 |
+
# (有無評価)
|
156 |
+
# 0は判断していない、又は判断が難しいという特徴量にする
|
157 |
+
hate = gr.State(0) # ヘイト
|
158 |
+
sexual = gr.State(0) # 性的
|
159 |
+
violence = gr.State(0) # 暴力
|
160 |
+
suicide = gr.State(0) # 自殺
|
161 |
+
threat = gr.State(0) # 犯罪
|
162 |
+
gun = gr.State(0) # 銃・重火器
|
163 |
+
controlled_substance = gr.State(0) # 規制対象物質
|
164 |
+
climinal_planing = gr.State(0) # 犯罪計画
|
165 |
+
privacy = gr.State(0) # 個人情報・プライバシー情報
|
166 |
+
harassment = gr.State(0) # ハラスメント
|
167 |
+
profanity = gr.State(0) # 冒涜
|
168 |
+
political_content = gr.State(0) # 政治的内容 (SteerLM)
|
169 |
+
moral_judgement = gr.State(0) # 非道徳 (SteerLM)
|
170 |
+
typos = gr.State(0) # 誤字脱字(minnadechat)
|
171 |
+
|
172 |
+
# 未整理 ========================================
|
173 |
+
|
174 |
+
# データセット読み込み
|
175 |
def dataset_load_fn() -> Tuple[
|
176 |
str,
|
177 |
str,
|
178 |
str,
|
179 |
str,
|
180 |
+
str,
|
181 |
+
str,
|
182 |
+
gr.update,
|
183 |
+
gr.update,
|
184 |
+
gr.update,
|
185 |
+
gr.update,
|
186 |
+
gr.update,
|
187 |
+
gr.update,
|
188 |
+
gr.update,
|
189 |
+
gr.update,
|
190 |
+
gr.update,
|
191 |
+
gr.update,
|
192 |
+
gr.update,
|
193 |
+
gr.update,
|
194 |
+
gr.update,
|
195 |
+
gr.update,
|
196 |
+
gr.update,
|
197 |
+
gr.update,
|
198 |
+
gr.update,
|
199 |
+
gr.update,
|
200 |
+
gr.update,
|
201 |
+
gr.update,
|
202 |
+
gr.update,
|
203 |
+
gr.update,
|
204 |
+
gr.update,
|
205 |
+
gr.update,
|
206 |
+
gr.update,
|
207 |
+
gr.update,
|
208 |
+
gr.update,
|
209 |
+
gr.update,
|
210 |
+
gr.update,
|
211 |
gr.update,
|
212 |
gr.update,
|
213 |
gr.update,
|
|
|
216 |
gr.update,
|
217 |
gr.update,
|
218 |
gr.update,
|
219 |
+
gr.update,
|
220 |
+
gr.update,
|
221 |
+
]:
|
222 |
|
223 |
+
is_dataset_loaded.value = False # ロード状態初期化
|
224 |
|
225 |
select_dataset.value = load_dataset(
|
226 |
select_dropdown_dataset.value
|
|
|
229 |
# DatasetオブジェクトをPandas DataFrameに変換
|
230 |
df = select_dataset.value["train"].to_pandas()
|
231 |
|
232 |
+
# ランダム化実施
|
233 |
# index列を追加し、シャッフル
|
234 |
df = df.reset_index(drop=False) # 元のindexを保持
|
235 |
df = df.sample(frac=1).reset_index(drop=True) # シャッフル
|
|
|
239 |
select_dataset_total_len.value = len(df) # 長さを取得
|
240 |
is_dataset_loaded.value = True # ロード完了
|
241 |
|
242 |
+
# データロード時にテキストの初期値を設定
|
243 |
initial_answer_text_1.value = df.iloc[select_idx.value]["answer"]
|
244 |
initial_answer_text_2.value = df.iloc[select_idx.value]["answer"]
|
245 |
|
|
|
247 |
df.iloc[select_idx.value]["answer"], \
|
248 |
df.iloc[select_idx.value]["question"], \
|
249 |
df.iloc[select_idx.value]["answer"], \
|
250 |
+
df.iloc[select_idx.value]["question"], \
|
251 |
+
df.iloc[select_idx.value]["answer"], \
|
252 |
+
gr.update(interactive=True), \
|
253 |
+
gr.update(interactive=True), \
|
254 |
+
gr.update(interactive=True), \
|
255 |
+
gr.update(interactive=True), \
|
256 |
+
gr.update(interactive=True), \
|
257 |
+
gr.update(interactive=True), \
|
258 |
+
gr.update(interactive=True), \
|
259 |
+
gr.update(interactive=True), \
|
260 |
+
gr.update(interactive=True), \
|
261 |
+
gr.update(interactive=True), \
|
262 |
+
gr.update(interactive=True), \
|
263 |
+
gr.update(interactive=True), \
|
264 |
+
gr.update(interactive=True), \
|
265 |
+
gr.update(interactive=True), \
|
266 |
+
gr.update(interactive=True), \
|
267 |
+
gr.update(interactive=True), \
|
268 |
+
gr.update(interactive=True), \
|
269 |
+
gr.update(interactive=True), \
|
270 |
+
gr.update(interactive=True), \
|
271 |
+
gr.update(interactive=True), \
|
272 |
+
gr.update(interactive=True), \
|
273 |
+
gr.update(interactive=True), \
|
274 |
+
gr.update(interactive=True), \
|
275 |
+
gr.update(interactive=True), \
|
276 |
+
gr.update(interactive=True), \
|
277 |
+
gr.update(interactive=True), \
|
278 |
+
gr.update(interactive=True), \
|
279 |
+
gr.update(interactive=True), \
|
280 |
+
gr.update(interactive=True), \
|
281 |
+
gr.update(interactive=True), \
|
282 |
+
gr.update(interactive=True), \
|
283 |
gr.update(interactive=True), \
|
284 |
gr.update(interactive=True), \
|
285 |
gr.update(interactive=True), \
|
|
|
298 |
annotated_folder = annotation_file.parent
|
299 |
|
300 |
scheduler = CommitScheduler(
|
301 |
+
repo_id=OUTPUT_DATASET[1],
|
302 |
repo_type="dataset",
|
303 |
folder_path=annotated_folder,
|
304 |
path_in_repo="data", # Spaceの場合の保存先フォルダー
|
|
|
306 |
every=5, # 5分毎にアップロード HuggingFAce_Documentの最低推奨値
|
307 |
)
|
308 |
|
309 |
+
# チェックボックスをintに変換
|
310 |
+
def checkbox_to_int(checkbox_value) -> int:
|
311 |
+
if checkbox_value == "不明":
|
312 |
+
output = 0
|
313 |
+
elif checkbox_value == "有":
|
314 |
+
output = 1
|
315 |
+
elif checkbox_value == "無":
|
316 |
+
output = 2
|
317 |
+
else :
|
318 |
+
output = 0
|
319 |
+
print("error: ")
|
320 |
+
return output
|
321 |
+
|
322 |
# CommitScheduler (HFへのデータアップロード
|
323 |
def save_annotation(
|
324 |
dataset_name: str,
|
325 |
dataset_id: int,
|
326 |
who: str,
|
327 |
+
unknown_quality: bool,
|
328 |
good: bool,
|
329 |
bad: bool,
|
|
|
|
|
330 |
is_proofreading_1: bool,
|
331 |
answer_text_1: str,
|
332 |
is_proofreading_2: bool,
|
333 |
+
answer_text_2: str,
|
334 |
+
score: int,
|
335 |
+
helphulness: int,
|
336 |
+
correcthness: int,
|
337 |
+
coherence: int,
|
338 |
+
complexity: int,
|
339 |
+
verbosity: int,
|
340 |
+
humor: int,
|
341 |
+
creativity: int,
|
342 |
+
appropriate: int,
|
343 |
+
following_instructions: int,
|
344 |
+
politeness: int,
|
345 |
+
harmfulness: int,
|
346 |
+
hate: int,
|
347 |
+
sexual: int,
|
348 |
+
violence: int,
|
349 |
+
suicide: int,
|
350 |
+
threat: int,
|
351 |
+
gun: int,
|
352 |
+
controlled_substance: int,
|
353 |
+
climinal_planing: int,
|
354 |
+
privacy: int,
|
355 |
+
harassment: int,
|
356 |
+
profanity: int,
|
357 |
+
political_content: int,
|
358 |
+
moral_judgement: int,
|
359 |
+
typos: int,
|
360 |
+
) -> None:
|
361 |
|
362 |
annotated_dataset.value = pd.concat([
|
363 |
annotated_dataset.value,
|
|
|
365 |
'dataset_name': [dataset_name],
|
366 |
'dataset_id': [dataset_id],
|
367 |
'who': [who],
|
368 |
+
'unknown_quality': [unknown_quality],
|
369 |
'good': [good],
|
370 |
'bad': [bad],
|
|
|
|
|
371 |
'is_proofreading_1': [is_proofreading_1],
|
372 |
"answer_text_1": [answer_text_1],
|
373 |
'is_proofreading_2': [is_proofreading_2],
|
374 |
+
'answer_text_2': [answer_text_2],
|
375 |
+
'score': [score],
|
376 |
+
'helphullness': [helphulness],
|
377 |
+
'correcthness': [correcthness],
|
378 |
+
'complexity': [complexity],
|
379 |
+
'verbosity': [verbosity],
|
380 |
+
'humor': [humor],
|
381 |
+
'creativity': [creativity],
|
382 |
+
'appropriate': [appropriate],
|
383 |
+
'following_finstructions': [following_instructions],
|
384 |
+
'politeness': [politeness],
|
385 |
+
'harmfulness': [harmfulness],
|
386 |
+
'hate': [hate],
|
387 |
+
'sextual': [sexual],
|
388 |
+
'violence': [violence],
|
389 |
+
'suicide': [suicide],
|
390 |
+
'threat': [threat],
|
391 |
+
'gun': [gun],
|
392 |
+
'controlled_substance': [controlled_substance],
|
393 |
+
'climinal_planing': [climinal_planing],
|
394 |
+
'privacy': [privacy],
|
395 |
+
'harassment': [harassment],
|
396 |
+
'profanity': [profanity],
|
397 |
+
'political_content': [political_content],
|
398 |
+
'moral_judgement': [moral_judgement],
|
399 |
+
'types': [typos]
|
400 |
})], ignore_index=True).reset_index(drop=True)
|
401 |
|
402 |
# 書き込み
|
403 |
with scheduler.lock:
|
404 |
with annotation_file.open("a", encoding='utf-8') as f:
|
405 |
data_to_write = {
|
406 |
+
# "id": , CommitSchedulerは、取得して末尾idを付与することが不可能?
|
407 |
"datetime": str(datetime.datetime.now().isoformat()),
|
408 |
"dataset_name": dataset_name,
|
409 |
"dataset_id": int(dataset_id),
|
410 |
"who": who,
|
411 |
+
"unknown_quality": unknown_quality,
|
412 |
"good": good,
|
413 |
"bad": bad,
|
|
|
|
|
414 |
"is_proofreading_1": is_proofreading_1,
|
415 |
"answer_text_1": answer_text_1,
|
416 |
"is_proofreading_2": is_proofreading_2,
|
417 |
"answer_text_2": answer_text_2,
|
418 |
+
"score": int(score),
|
419 |
+
"helphulness": int(helphulness),
|
420 |
+
"correcthness": int(correcthness),
|
421 |
+
"coherence": int(coherence),
|
422 |
+
"complexity": int(complexity),
|
423 |
+
"verbosity": int(verbosity),
|
424 |
+
"humor": int(humor),
|
425 |
+
"creativity": int(creativity),
|
426 |
+
"appropriate": int(appropriate),
|
427 |
+
"following_instructions": int(following_instructions),
|
428 |
+
"politeness": int(politeness),
|
429 |
+
"harmfulness": int(harmfulness),
|
430 |
+
"hate": int(hate),
|
431 |
+
"sexual": int(sexual),
|
432 |
+
"violence": int(violence),
|
433 |
+
"suicide": int(suicide),
|
434 |
+
"threat": int(threat),
|
435 |
+
"gun": int(gun),
|
436 |
+
"controlled_substance": int(controlled_substance),
|
437 |
+
"climinal_planing": int(climinal_planing),
|
438 |
+
"privacy": int(privacy),
|
439 |
+
"harassment": int(harassment),
|
440 |
+
"profanity": int(profanity),
|
441 |
+
"policital_content": int(political_content),
|
442 |
+
"moral_judgement": int(moral_judgement),
|
443 |
+
"typos": int(typos)
|
444 |
}
|
445 |
f.write(json.dumps(data_to_write, ensure_ascii=False))
|
446 |
f.write("\n")
|
|
|
452 |
|
453 |
# ユーザー名表示
|
454 |
|
|
|
455 |
def hello(profile: gr.OAuthProfile | None) -> Tuple[str, str]:
|
456 |
if profile is None:
|
457 |
return "プライベートデータセット取得のためにログインしてください。", who.value
|
|
|
463 |
theme_ = gr.themes.Default()
|
464 |
|
465 |
# 後のCSSデザイン変更用
|
|
|
|
|
466 |
def load_css():
|
467 |
with open("style.css", "r") as file:
|
468 |
css_content = file.read()
|
469 |
return css_content
|
470 |
|
|
|
471 |
# Gradio 画面 ============================================
|
472 |
with gr.Blocks(theme=theme_, css=load_css()) as demo:
|
473 |
|
474 |
gr.Markdown("# データセット アノテーション for Tanuki (Phase2)")
|
475 |
+
|
476 |
+
with gr.Row():
|
477 |
+
gr.Markdown("### GENIACにて開発中のLLM用データセットに対してアノテーションするSpaceです\n \
|
478 |
+
入力されたデータは使用される可能性があるため、個人情報・秘匿情報などは入力しないでください。\n \
|
479 |
+
テスト中です。")
|
480 |
+
gr_who = gr.TextArea(value=who.value, lines=1, max_lines=1, label="ユーザー名 (入力してください)")
|
481 |
+
|
482 |
+
def change_name(name: str):
|
483 |
+
who.value = name
|
484 |
+
|
485 |
+
gr_who.change(
|
486 |
+
change_name,
|
487 |
+
inputs=[gr_who],
|
488 |
+
outputs=[]
|
489 |
+
)
|
490 |
|
491 |
with gr.Tab("アノテーション (シングルターン)"):
|
492 |
|
493 |
+
# HF login 機能
|
494 |
+
# with gr.Row(equal_height=True):
|
495 |
+
# gr.LoginButton(value="HuggingFace ログイン",
|
496 |
+
# logout_value="HuggingFace ログアウト", scale=1)
|
|
|
497 |
# ユーザー名
|
498 |
+
# gr_profile_name = gr.Markdown()
|
499 |
+
# demo.load(hello, inputs=None, outputs=[gr_profile_name, who])
|
500 |
|
501 |
with gr.Row():
|
502 |
|
|
|
519 |
gr_data_load_btn = gr.Button("② データセットを読み込む")
|
520 |
|
521 |
with gr.Column() as content_column:
|
522 |
+
|
523 |
with gr.Tab("③ シンプル(良・悪)"):
|
524 |
with gr.Column():
|
525 |
with gr.Row(equal_height=True):
|
526 |
+
gr_good_btn = gr.Button("良い", interactive=False)
|
527 |
+
gr_unknown_btn = gr.Button(
|
528 |
"分からない", interactive=False) # 「分からない」ボタンを追加
|
529 |
+
gr_bad_btn = gr.Button("悪い", interactive=False)
|
530 |
|
531 |
+
gr_question_text_1_1 = gr.Textbox(
|
532 |
label="質問: ", lines=5, interactive=False)
|
533 |
|
534 |
+
gr_answer_text_1_1 = gr.Textbox(
|
535 |
+
label="回答: 訂正して頂けると、品質が上がります。",
|
536 |
lines=20,
|
537 |
+
interactive=False)
|
538 |
+
|
539 |
+
with gr.Tab("③ 5段階評価(シンプル)"):
|
540 |
+
|
541 |
+
with gr.Row() as simple_score_btn:
|
542 |
+
gr_score_5_btn = gr.Button("5: 高品質", interactive=False)
|
543 |
+
gr_score_4_btn = gr.Button("4: 良い", interactive=False)
|
544 |
+
gr_score_3_btn = gr.Button("3: 普通", interactive=False)
|
545 |
+
gr_score_2_btn = gr.Button("2: 悪い", interactive=False)
|
546 |
+
gr_score_1_btn = gr.Button("1: 低品質", interactive=False)
|
547 |
+
|
548 |
+
gr_question_text_2_1 = gr.Textbox(
|
549 |
+
label="質問: ", lines=5, interactive=False)
|
550 |
+
|
551 |
+
gr_answer_text_2_1 = gr.Textbox(
|
552 |
+
label="回答: 訂正して頂けると、品質が上がります。", lines=20, interactive=False)
|
553 |
+
|
554 |
+
with gr.Tab("③ 5段階評価 (詳細)"):
|
555 |
+
|
556 |
+
with gr.Row():
|
557 |
+
|
558 |
+
with gr.Column() as EvalFive:
|
559 |
+
|
560 |
+
gr_question_text_3_1 = gr.Textbox(
|
561 |
+
label="質問: ", lines=10, interactive=False)
|
562 |
+
|
563 |
+
# gr_explain = gr.Markdown("回答")
|
564 |
+
|
565 |
+
gr_answer_text_3_1 = gr.Textbox(
|
566 |
+
label="回答: 訂正して頂けると、品質が上がります。", lines=35, interactive=False)
|
567 |
+
|
568 |
+
with gr.Column() as EvalFiveDetail:
|
569 |
+
|
570 |
+
with gr.Row():
|
571 |
+
gr_submit_score = gr.Button("評価送信", interactive=False)
|
572 |
+
gr_score_reset = gr.Button("スコアリセット", interactive=False)
|
573 |
+
|
574 |
+
gr_score_detail = gr.Slider(label="総合スコア 【必須】",value=3, minimum=1, maximum=5, step=1, interactive=False)
|
575 |
+
gr_evalannotation_explain = gr.Markdown("詳細アノテーション (5点満点)")
|
576 |
+
|
577 |
+
with gr.Row() :
|
578 |
+
gr_helphulness = gr.Slider(label="有用性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
579 |
+
gr_correcthness = gr.Slider(label="正確さ", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
580 |
+
gr_coherence = gr.Slider(label="一貫性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
581 |
+
gr_complexity = gr.Slider(label="複雑さ", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
582 |
+
|
583 |
+
with gr.Row() :
|
584 |
+
gr_verbosity = gr.Slider(label="冗長性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
585 |
+
gr_humor = gr.Slider(label="ユーモア", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
586 |
+
gr_creativity = gr.Slider(label="創造性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
587 |
+
gr_appropriate = gr.Slider(label="適切性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
588 |
+
|
589 |
+
with gr.Row() :
|
590 |
+
gr_following_instructions = gr.Slider(label="忠実性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
591 |
+
gr_politeness = gr.Slider(label="礼儀正しさ", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
592 |
+
gr_harmfulness = gr.Slider(label="有害性", minimum=0, value=0, maximum=5, step=1, interactive=False)
|
593 |
+
gr_text_4 = gr.Markdown("判断可能な物のみ\n\nスコアを付けてください。")
|
594 |
+
|
595 |
+
gr_text_5 = gr.Markdown("")
|
596 |
+
|
597 |
+
with gr.Row():
|
598 |
+
gr_typos = gr.Radio(label="誤字・脱字", choices=["不明", "有", "無"], value="不明", interactive=False)
|
599 |
+
gr_moral_judgement = gr.Radio(label="非道徳", choices=["不明", "有", "無"], value="不明", interactive=False)
|
600 |
+
|
601 |
+
with gr.Row():
|
602 |
+
gr_hate = gr.Radio(label="ヘイト", choices=["不明", "有", "無"], value="不明", interactive=False)
|
603 |
+
gr_sexual = gr.Radio(label="性的内容", choices=["不明", "有", "無"], value="不明", interactive=False)
|
604 |
+
|
605 |
+
with gr.Row():
|
606 |
+
gr_violence = gr.Radio(label="暴力的", choices=["不明", "有", "無"], value="不明", interactive=False)
|
607 |
+
gr_suicide = gr.Radio(label="自殺行為", choices=["不明", "有", "無"], value="不明", interactive=False)
|
608 |
+
|
609 |
+
with gr.Row():
|
610 |
+
gr_threat = gr.Radio(label="犯罪", choices=["不明", "有", "無"], value="不明", interactive=False)
|
611 |
+
gr_gun = gr.Radio(label="銃等", choices=["不明", "有", "無"], value="不明", interactive=False)
|
612 |
+
|
613 |
+
with gr.Row():
|
614 |
+
gr_controlled_substance = gr.Radio(label="規制対象物質", choices=["不明", "有", "無"], value="不明", interactive=False)
|
615 |
+
gr_climinal_planing = gr.Radio(label="犯罪計画", choices=["不明", "有", "無"], value="不明", interactive=False)
|
616 |
+
|
617 |
+
with gr.Row():
|
618 |
+
gr_privacy = gr.Radio(label="個人情報", choices=["不明", "有", "無"], value="不明", interactive=False)
|
619 |
+
gr_harassment = gr.Radio(label="ハラスメント", choices=["不明", "有", "無"], value="不明", interactive=False)
|
620 |
+
|
621 |
+
with gr.Row():
|
622 |
+
gr_profanity = gr.Radio(label="冒涜行為", choices=["不明", "有", "無"], value="不明", interactive=False)
|
623 |
+
gr_political_content = gr.Radio(label="政治的内容", choices=["不明", "有", "無"], value="不明", interactive=False)
|
624 |
+
|
625 |
+
|
626 |
+
# 5段階評価 詳細のgr.State Update
|
627 |
+
def eval_submit(
|
628 |
+
# text
|
629 |
+
g_answer_text_3_1: str,
|
630 |
+
|
631 |
+
# slicer
|
632 |
+
g_score: int,
|
633 |
+
g_helphulness: int,
|
634 |
+
g_correcthness: int,
|
635 |
+
g_coherenece: int,
|
636 |
+
g_complexity: int,
|
637 |
+
g_verbosity: int,
|
638 |
+
g_humor: int,
|
639 |
+
g_creativity: int,
|
640 |
+
g_appropriate: int,
|
641 |
+
g_following_instructions: int,
|
642 |
+
g_politeness: int,
|
643 |
+
g_harmfulness: int,
|
644 |
+
|
645 |
+
# checkbox
|
646 |
+
g_hate: str,
|
647 |
+
g_sexual: str,
|
648 |
+
g_violence: str,
|
649 |
+
g_suicide: str,
|
650 |
+
g_threat: str,
|
651 |
+
g_gun: str,
|
652 |
+
g_controlled_substance: str,
|
653 |
+
g_climinal_planing: str,
|
654 |
+
g_privacy: str,
|
655 |
+
g_harassment: str,
|
656 |
+
g_profanity: str,
|
657 |
+
g_political_content: str,
|
658 |
+
g_moral_judgement: str,
|
659 |
+
g_typos: str,
|
660 |
+
):
|
661 |
+
# 状態初期化
|
662 |
+
good.value = False
|
663 |
+
bad.value = False
|
664 |
+
unknown_quality.value = False
|
665 |
+
|
666 |
+
# テキストに更新があったかどうかチェック
|
667 |
+
if initial_answer_text_1.value != g_answer_text_3_1:
|
668 |
+
is_proofreading_1.value = True
|
669 |
+
answer_text_1.value = g_answer_text_3_1
|
670 |
+
else:
|
671 |
+
is_proofreading_1.value = False
|
672 |
+
answer_text_1.value = ""
|
673 |
+
|
674 |
+
# 送信用データ更新
|
675 |
+
score.value = g_score
|
676 |
+
helphulness.value = g_helphulness
|
677 |
+
correcthness.value = g_correcthness
|
678 |
+
coherenece.value = g_coherenece
|
679 |
+
complexity.value = g_complexity
|
680 |
+
verbosity.value = g_verbosity
|
681 |
+
humor.value = g_humor
|
682 |
+
creativity.value = g_creativity
|
683 |
+
appropriate.value = g_appropriate
|
684 |
+
following_instructions.value = g_following_instructions
|
685 |
+
politeness.value = g_politeness
|
686 |
+
harmfulness.value = g_harmfulness
|
687 |
+
|
688 |
+
hate.value = checkbox_to_int(g_hate)
|
689 |
+
sexual.value = checkbox_to_int(g_sexual)
|
690 |
+
violence.value = checkbox_to_int(g_violence)
|
691 |
+
suicide.value = checkbox_to_int(g_suicide)
|
692 |
+
threat.value = checkbox_to_int(g_threat)
|
693 |
+
gun.value = checkbox_to_int(g_gun)
|
694 |
+
controlled_substance.value = checkbox_to_int(g_controlled_substance)
|
695 |
+
climinal_planing.value = checkbox_to_int(g_climinal_planing)
|
696 |
+
privacy.value = checkbox_to_int(g_privacy)
|
697 |
+
harassment.value = checkbox_to_int(g_harassment)
|
698 |
+
profanity.value = checkbox_to_int(g_profanity)
|
699 |
+
political_content.value = checkbox_to_int(g_political_content)
|
700 |
+
moral_judgement.value = checkbox_to_int(g_moral_judgement)
|
701 |
+
typos.value = checkbox_to_int(g_typos)
|
702 |
+
|
703 |
+
# データ送信
|
704 |
+
|
705 |
+
# 表示更新
|
706 |
+
# indexを進める
|
707 |
+
select_idx.value += 1
|
708 |
|
709 |
+
df = select_dataset.value["train"]
|
710 |
|
711 |
+
# ループさせるか、エラー処理を行う
|
712 |
+
if select_idx.value >= len(df):
|
713 |
+
select_idx.value = 0
|
714 |
|
715 |
+
# データセットに追加
|
716 |
+
df = select_dataset.value["train"]
|
717 |
+
save_annotation(
|
718 |
+
select_dropdown_dataset.value,
|
719 |
+
df.iloc[select_idx.value]['index'],
|
720 |
+
who.value,
|
721 |
+
unknown_quality.value,
|
722 |
+
good.value,
|
723 |
+
bad.value,
|
724 |
+
is_proofreading_1.value,
|
725 |
+
answer_text_1.value,
|
726 |
+
is_proofreading_2.value,
|
727 |
+
answer_text_2.value,
|
728 |
+
score.value,
|
729 |
+
helphulness.value,
|
730 |
+
correcthness.value,
|
731 |
+
coherenece.value,
|
732 |
+
complexity.value,
|
733 |
+
verbosity.value,
|
734 |
+
humor.value,
|
735 |
+
creativity.value,
|
736 |
+
appropriate.value,
|
737 |
+
following_instructions.value,
|
738 |
+
politeness.value,
|
739 |
+
harmfulness.value,
|
740 |
+
hate.value,
|
741 |
+
sexual.value,
|
742 |
+
violence.value,
|
743 |
+
suicide.value,
|
744 |
+
threat.value,
|
745 |
+
gun.value,
|
746 |
+
controlled_substance.value,
|
747 |
+
climinal_planing.value,
|
748 |
+
privacy.value,
|
749 |
+
harassment.value,
|
750 |
+
profanity.value,
|
751 |
+
political_content.value,
|
752 |
+
moral_judgement.value,
|
753 |
+
typos.value,
|
754 |
+
)
|
755 |
|
756 |
+
# Nextデータ初期化
|
757 |
+
initialize_next_data(df)
|
758 |
+
|
759 |
+
# 5段階評価をリセット
|
760 |
+
score_reset_display()
|
761 |
+
|
762 |
+
# UI更新のための返り値
|
763 |
+
return gr.update(value=df.iloc[select_idx.value]["question"]), \
|
764 |
+
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
765 |
+
gr.update(value=df.iloc[select_idx.value]["question"]), \
|
766 |
+
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
767 |
+
gr.update(value=df.iloc[select_idx.value]["question"]), \
|
768 |
+
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
769 |
+
*score_reset_display() # スコアリセット用の返り値
|
770 |
+
|
771 |
+
gr_submit_score.click(
|
772 |
+
eval_submit,
|
773 |
+
inputs=[
|
774 |
+
gr_answer_text_3_1,
|
775 |
+
gr_score_detail,
|
776 |
+
gr_helphulness,
|
777 |
+
gr_correcthness,
|
778 |
+
gr_coherence,
|
779 |
+
gr_complexity,
|
780 |
+
gr_verbosity,
|
781 |
+
gr_humor,
|
782 |
+
gr_creativity,
|
783 |
+
gr_appropriate,
|
784 |
+
gr_following_instructions,
|
785 |
+
gr_politeness,
|
786 |
+
gr_harmfulness,
|
787 |
+
gr_hate,
|
788 |
+
gr_sexual,
|
789 |
+
gr_violence,
|
790 |
+
gr_suicide,
|
791 |
+
gr_threat,
|
792 |
+
gr_gun,
|
793 |
+
gr_controlled_substance,
|
794 |
+
gr_climinal_planing,
|
795 |
+
gr_privacy,
|
796 |
+
gr_harassment,
|
797 |
+
gr_profanity,
|
798 |
+
gr_political_content,
|
799 |
+
gr_moral_judgement,
|
800 |
+
gr_typos
|
801 |
+
],
|
802 |
+
# UI更新
|
803 |
+
outputs=[
|
804 |
+
gr_question_text_1_1,
|
805 |
+
gr_answer_text_1_1,
|
806 |
+
gr_question_text_2_1,
|
807 |
+
gr_answer_text_2_1,
|
808 |
+
gr_question_text_3_1,
|
809 |
+
gr_answer_text_3_1,
|
810 |
+
gr_score_detail,
|
811 |
+
gr_helphulness,
|
812 |
+
gr_correcthness,
|
813 |
+
gr_coherence,
|
814 |
+
gr_complexity,
|
815 |
+
gr_verbosity,
|
816 |
+
gr_humor,
|
817 |
+
gr_creativity,
|
818 |
+
gr_appropriate,
|
819 |
+
gr_following_instructions,
|
820 |
+
gr_politeness,
|
821 |
+
gr_harmfulness,
|
822 |
+
gr_hate,
|
823 |
+
gr_sexual,
|
824 |
+
gr_violence,
|
825 |
+
gr_suicide,
|
826 |
+
gr_threat,
|
827 |
+
gr_gun,
|
828 |
+
gr_controlled_substance,
|
829 |
+
gr_climinal_planing,
|
830 |
+
gr_privacy,
|
831 |
+
gr_harassment,
|
832 |
+
gr_profanity,
|
833 |
+
gr_political_content,
|
834 |
+
gr_moral_judgement,
|
835 |
+
gr_typos
|
836 |
+
],
|
837 |
+
)
|
838 |
|
839 |
# 5段階評価ボタンのクリックイベントを定義
|
840 |
def score_button_clicked(button_value):
|
841 |
+
# 状態初期化
|
842 |
good.value = False
|
843 |
bad.value = False
|
844 |
+
unknown_quality.value = False
|
845 |
+
score.value = button_value
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
846 |
|
847 |
+
# 5段階評価クリック
|
848 |
def update_annotation(
|
849 |
input_ans_1: str = None,
|
850 |
+
input_ans_2: str = None,
|
851 |
is_good: bool = None, # good/bad を表すフラグを追加
|
852 |
is_unknown: bool = None, # 「分からない」を表すフラグを追加
|
853 |
+
is_simple: bool = None,
|
854 |
score_value: int = None # 5段階評価の値、good/badの場合はNone
|
855 |
+
) -> Tuple[
|
856 |
+
gr.update,
|
857 |
+
gr.update,
|
858 |
+
gr.update,
|
859 |
+
gr.update,
|
860 |
+
gr.update,
|
861 |
+
]:
|
862 |
|
863 |
# good/bad/unknown と score の状態を更新
|
864 |
update_evaluation_state(is_good, is_unknown, score_value)
|
865 |
|
866 |
# 変更を検知 (5段階評価の場合も処理するように変更)
|
867 |
+
update_answer_state(input_ans_1, input_ans_2)
|
868 |
+
|
869 |
# 表示更新
|
870 |
# indexを進める
|
871 |
select_idx.value += 1
|
|
|
875 |
# ループさせるか、エラー処理を行う
|
876 |
if select_idx.value >= len(df):
|
877 |
select_idx.value = 0
|
878 |
+
|
879 |
+
# シンプルモードクリックで、scoreを0に設定
|
880 |
+
if is_simple == True:
|
881 |
+
score.value = 0
|
882 |
|
883 |
# データセットに追加
|
884 |
+
# 元のindex番号(dataset_id)を指定して保存
|
885 |
save_annotation_data(df)
|
886 |
|
887 |
# Nextデータ初期化
|
888 |
initialize_next_data(df)
|
889 |
+
|
890 |
+
# 5段階評価をリセット
|
891 |
+
reset_values = score_reset_display()
|
892 |
|
893 |
return gr.update(value=df.iloc[select_idx.value]["question"]), \
|
894 |
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
895 |
gr.update(value=df.iloc[select_idx.value]["question"]), \
|
896 |
+
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
897 |
+
gr.update(value=df.iloc[select_idx.value]["question"]), \
|
898 |
+
gr.update(value=df.iloc[select_idx.value]["answer"]), \
|
899 |
+
*reset_values
|
900 |
+
|
901 |
|
902 |
+
# 評価状態リセット
|
903 |
def update_evaluation_state(is_good, is_unknown, score_value):
|
904 |
if score_value is not None: # 5段階評価の場合
|
905 |
good.value = False
|
|
|
911 |
bad.value = not is_good if not is_unknown else False
|
912 |
unknown_quality.value = is_unknown # 「分からない」の状態を設定
|
913 |
|
914 |
+
# 変更を検知して値を設定
|
915 |
+
def update_answer_state(input_ans_1, input_ans_2):
|
916 |
if input_ans_1 is not None and initial_answer_text_1.value != input_ans_1:
|
917 |
is_proofreading_1.value = True
|
918 |
answer_text_1.value = input_ans_1
|
919 |
else:
|
920 |
answer_text_1.value = ""
|
921 |
|
922 |
+
# 2ターン目用
|
923 |
+
if input_ans_2 is not None and initial_answer_text_2.value != input_ans_2:
|
924 |
is_proofreading_2.value = True
|
925 |
+
answer_text_2.value = input_ans_2
|
926 |
else:
|
927 |
answer_text_2.value = ""
|
928 |
|
929 |
+
# データ保存
|
930 |
def save_annotation_data(df):
|
931 |
save_annotation(
|
932 |
select_dropdown_dataset.value,
|
933 |
# datasetIdは元のindex番号を使用
|
934 |
df.iloc[select_idx.value]['index'],
|
935 |
who.value,
|
936 |
+
unknown_quality.value,
|
937 |
good.value,
|
938 |
bad.value,
|
|
|
|
|
939 |
is_proofreading_1.value,
|
940 |
answer_text_1.value,
|
941 |
is_proofreading_2.value,
|
942 |
+
answer_text_2.value,
|
943 |
+
score.value,
|
944 |
+
helphulness.value,
|
945 |
+
correcthness.value,
|
946 |
+
coherenece.value,
|
947 |
+
complexity.value,
|
948 |
+
verbosity.value,
|
949 |
+
humor.value,
|
950 |
+
creativity.value,
|
951 |
+
appropriate.value,
|
952 |
+
following_instructions.value,
|
953 |
+
politeness.value,
|
954 |
+
harmfulness.value,
|
955 |
+
hate.value,
|
956 |
+
sexual.value,
|
957 |
+
violence.value,
|
958 |
+
suicide.value,
|
959 |
+
threat.value,
|
960 |
+
gun.value,
|
961 |
+
controlled_substance.value,
|
962 |
+
climinal_planing.value,
|
963 |
+
privacy.value,
|
964 |
+
harassment.value,
|
965 |
+
profanity.value,
|
966 |
+
political_content.value,
|
967 |
+
moral_judgement.value,
|
968 |
+
typos.value,
|
969 |
)
|
970 |
|
971 |
+
# 次データ読込時の初期化
|
972 |
def initialize_next_data(df):
|
973 |
is_proofreading_1.value = False
|
974 |
is_proofreading_2.value = False
|
975 |
initial_answer_text_1.value = df.iloc[select_idx.value]["answer"]
|
976 |
initial_answer_text_2.value = df.iloc[select_idx.value]["answer"]
|
977 |
+
|
978 |
+
# 評価不可
|
979 |
+
def unknown_click(input_ans_1, input_ans_2):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
980 |
good.value = False
|
981 |
bad.value = False
|
982 |
+
score_reset_display()
|
983 |
+
return update_annotation(
|
984 |
+
input_ans_1=input_ans_1,
|
985 |
+
input_ans_2=input_ans_2,
|
986 |
+
is_good=False,
|
987 |
+
is_unknown=True,
|
988 |
+
is_simple=True,
|
989 |
+
)
|
990 |
+
|
991 |
+
# 評価不可クリックイベント
|
992 |
+
gr_unknown_btn.click(
|
993 |
unknown_click,
|
994 |
inputs=[
|
995 |
+
gr_answer_text_1_1,
|
996 |
+
gr_answer_text_2_1
|
997 |
],
|
998 |
+
outputs=[
|
999 |
+
gr_question_text_1_1,
|
1000 |
+
gr_answer_text_1_1,
|
1001 |
+
gr_question_text_2_1,
|
1002 |
+
gr_answer_text_2_1,
|
1003 |
+
gr_question_text_3_1,
|
1004 |
+
gr_answer_text_3_1,
|
1005 |
+
gr_score_detail,
|
1006 |
+
gr_helphulness,
|
1007 |
+
gr_correcthness,
|
1008 |
+
gr_coherence,
|
1009 |
+
gr_complexity,
|
1010 |
+
gr_verbosity,
|
1011 |
+
gr_humor,
|
1012 |
+
gr_creativity,
|
1013 |
+
gr_appropriate,
|
1014 |
+
gr_following_instructions,
|
1015 |
+
gr_politeness,
|
1016 |
+
gr_harmfulness,
|
1017 |
+
gr_hate,
|
1018 |
+
gr_sexual,
|
1019 |
+
gr_violence,
|
1020 |
+
gr_suicide,
|
1021 |
+
gr_threat,
|
1022 |
+
gr_gun,
|
1023 |
+
gr_controlled_substance,
|
1024 |
+
gr_climinal_planing,
|
1025 |
+
gr_privacy,
|
1026 |
+
gr_harassment,
|
1027 |
+
gr_profanity,
|
1028 |
+
gr_political_content,
|
1029 |
+
gr_moral_judgement,
|
1030 |
+
gr_typos
|
1031 |
+
]
|
1032 |
)
|
1033 |
|
1034 |
+
# 良いクリック
|
1035 |
+
def good_click(input_ans_1, input_ans_2):
|
1036 |
+
score_reset_display()
|
1037 |
+
return update_annotation(
|
1038 |
+
input_ans_1=input_ans_1,
|
1039 |
+
input_ans_2=input_ans_2,
|
1040 |
+
is_good=True,
|
1041 |
+
is_unknown=False,
|
1042 |
+
is_simple=True,
|
1043 |
+
)
|
1044 |
+
|
1045 |
+
# 良いクリック
|
1046 |
+
gr_good_btn.click(
|
1047 |
+
good_click,
|
1048 |
+
inputs=[
|
1049 |
+
gr_answer_text_1_1,
|
1050 |
+
gr_answer_text_2_1
|
1051 |
+
],
|
1052 |
+
outputs=[
|
1053 |
+
gr_question_text_1_1,
|
1054 |
+
gr_answer_text_1_1,
|
1055 |
+
gr_question_text_2_1,
|
1056 |
+
gr_answer_text_2_1,
|
1057 |
+
gr_question_text_3_1,
|
1058 |
+
gr_answer_text_3_1,
|
1059 |
+
gr_score_detail,
|
1060 |
+
gr_helphulness,
|
1061 |
+
gr_correcthness,
|
1062 |
+
gr_coherence,
|
1063 |
+
gr_complexity,
|
1064 |
+
gr_verbosity,
|
1065 |
+
gr_humor,
|
1066 |
+
gr_creativity,
|
1067 |
+
gr_appropriate,
|
1068 |
+
gr_following_instructions,
|
1069 |
+
gr_politeness,
|
1070 |
+
gr_harmfulness,
|
1071 |
+
gr_hate,
|
1072 |
+
gr_sexual,
|
1073 |
+
gr_violence,
|
1074 |
+
gr_suicide,
|
1075 |
+
gr_threat,
|
1076 |
+
gr_gun,
|
1077 |
+
gr_controlled_substance,
|
1078 |
+
gr_climinal_planing,
|
1079 |
+
gr_privacy,
|
1080 |
+
gr_harassment,
|
1081 |
+
gr_profanity,
|
1082 |
+
gr_political_content,
|
1083 |
+
gr_moral_judgement,
|
1084 |
+
gr_typos
|
1085 |
+
]
|
1086 |
+
)
|
1087 |
+
|
1088 |
+
# 低評価クリック
|
1089 |
+
def bad_click(input_ans_1, input_ans_2):
|
1090 |
+
score_reset_display()
|
1091 |
+
return update_annotation(
|
1092 |
+
input_ans_1=input_ans_1,
|
1093 |
+
input_ans_2=input_ans_2,
|
1094 |
+
is_good=False,
|
1095 |
+
is_unknown=False,
|
1096 |
+
is_simple=True,
|
1097 |
+
)
|
1098 |
+
|
1099 |
+
# 低評価クリックイベント
|
1100 |
+
gr_bad_btn.click(
|
1101 |
bad_click,
|
1102 |
inputs=[
|
1103 |
+
gr_answer_text_1_1,
|
1104 |
+
gr_answer_text_2_1
|
1105 |
],
|
1106 |
+
outputs=[
|
1107 |
+
gr_question_text_1_1,
|
1108 |
+
gr_answer_text_1_1,
|
1109 |
+
gr_question_text_2_1,
|
1110 |
+
gr_answer_text_2_1,
|
1111 |
+
gr_question_text_3_1,
|
1112 |
+
gr_answer_text_3_1,
|
1113 |
+
gr_score_detail,
|
1114 |
+
gr_helphulness,
|
1115 |
+
gr_correcthness,
|
1116 |
+
gr_coherence,
|
1117 |
+
gr_complexity,
|
1118 |
+
gr_verbosity,
|
1119 |
+
gr_humor,
|
1120 |
+
gr_creativity,
|
1121 |
+
gr_appropriate,
|
1122 |
+
gr_following_instructions,
|
1123 |
+
gr_politeness,
|
1124 |
+
gr_harmfulness,
|
1125 |
+
gr_hate,
|
1126 |
+
gr_sexual,
|
1127 |
+
gr_violence,
|
1128 |
+
gr_suicide,
|
1129 |
+
gr_threat,
|
1130 |
+
gr_gun,
|
1131 |
+
gr_controlled_substance,
|
1132 |
+
gr_climinal_planing,
|
1133 |
+
gr_privacy,
|
1134 |
+
gr_harassment,
|
1135 |
+
gr_profanity,
|
1136 |
+
gr_political_content,
|
1137 |
+
gr_moral_judgement,
|
1138 |
+
gr_typos
|
1139 |
+
]
|
1140 |
)
|
1141 |
|
1142 |
# 5段階評価ボタンのクリックイベント
|
1143 |
+
gr_score_1_btn.click(lambda x: update_annotation(input_ans_1=x, is_unknown=False, score_value=1),
|
1144 |
+
inputs=[gr_answer_text_2_1],
|
1145 |
+
outputs=[
|
1146 |
+
gr_question_text_1_1,
|
1147 |
+
gr_answer_text_1_1,
|
1148 |
+
gr_question_text_2_1,
|
1149 |
+
gr_answer_text_2_1,
|
1150 |
+
gr_question_text_3_1,
|
1151 |
+
gr_answer_text_3_1,
|
1152 |
+
gr_score_detail,
|
1153 |
+
gr_helphulness,
|
1154 |
+
gr_correcthness,
|
1155 |
+
gr_coherence,
|
1156 |
+
gr_complexity,
|
1157 |
+
gr_verbosity,
|
1158 |
+
gr_humor,
|
1159 |
+
gr_creativity,
|
1160 |
+
gr_appropriate,
|
1161 |
+
gr_following_instructions,
|
1162 |
+
gr_politeness,
|
1163 |
+
gr_harmfulness,
|
1164 |
+
gr_hate,
|
1165 |
+
gr_sexual,
|
1166 |
+
gr_violence,
|
1167 |
+
gr_suicide,
|
1168 |
+
gr_threat,
|
1169 |
+
gr_gun,
|
1170 |
+
gr_controlled_substance,
|
1171 |
+
gr_climinal_planing,
|
1172 |
+
gr_privacy,
|
1173 |
+
gr_harassment,
|
1174 |
+
gr_profanity,
|
1175 |
+
gr_political_content,
|
1176 |
+
gr_moral_judgement,
|
1177 |
+
gr_typos
|
1178 |
+
])
|
1179 |
+
gr_score_2_btn.click(lambda x: update_annotation(input_ans_1=x, is_unknown=False, score_value=2),
|
1180 |
+
inputs=[gr_answer_text_2_1],
|
1181 |
+
outputs=[
|
1182 |
+
gr_question_text_1_1,
|
1183 |
+
gr_answer_text_1_1,
|
1184 |
+
gr_question_text_2_1,
|
1185 |
+
gr_answer_text_2_1,
|
1186 |
+
gr_question_text_3_1,
|
1187 |
+
gr_answer_text_3_1,
|
1188 |
+
gr_score_detail,
|
1189 |
+
gr_helphulness,
|
1190 |
+
gr_correcthness,
|
1191 |
+
gr_coherence,
|
1192 |
+
gr_complexity,
|
1193 |
+
gr_verbosity,
|
1194 |
+
gr_humor,
|
1195 |
+
gr_creativity,
|
1196 |
+
gr_appropriate,
|
1197 |
+
gr_following_instructions,
|
1198 |
+
gr_politeness,
|
1199 |
+
gr_harmfulness,
|
1200 |
+
gr_hate,
|
1201 |
+
gr_sexual,
|
1202 |
+
gr_violence,
|
1203 |
+
gr_suicide,
|
1204 |
+
gr_threat,
|
1205 |
+
gr_gun,
|
1206 |
+
gr_controlled_substance,
|
1207 |
+
gr_climinal_planing,
|
1208 |
+
gr_privacy,
|
1209 |
+
gr_harassment,
|
1210 |
+
gr_profanity,
|
1211 |
+
gr_political_content,
|
1212 |
+
gr_moral_judgement,
|
1213 |
+
gr_typos
|
1214 |
+
])
|
1215 |
+
gr_score_3_btn.click(lambda x: update_annotation(input_ans_1=x, is_unknown=False, score_value=3),
|
1216 |
+
inputs=[gr_answer_text_2_1],
|
1217 |
+
outputs=[
|
1218 |
+
gr_question_text_1_1,
|
1219 |
+
gr_answer_text_1_1,
|
1220 |
+
gr_question_text_2_1,
|
1221 |
+
gr_answer_text_2_1,
|
1222 |
+
gr_question_text_3_1,
|
1223 |
+
gr_answer_text_3_1,
|
1224 |
+
gr_score_detail,
|
1225 |
+
gr_helphulness,
|
1226 |
+
gr_correcthness,
|
1227 |
+
gr_coherence,
|
1228 |
+
gr_complexity,
|
1229 |
+
gr_verbosity,
|
1230 |
+
gr_humor,
|
1231 |
+
gr_creativity,
|
1232 |
+
gr_appropriate,
|
1233 |
+
gr_following_instructions,
|
1234 |
+
gr_politeness,
|
1235 |
+
gr_harmfulness,
|
1236 |
+
gr_hate,
|
1237 |
+
gr_sexual,
|
1238 |
+
gr_violence,
|
1239 |
+
gr_suicide,
|
1240 |
+
gr_threat,
|
1241 |
+
gr_gun,
|
1242 |
+
gr_controlled_substance,
|
1243 |
+
gr_climinal_planing,
|
1244 |
+
gr_privacy,
|
1245 |
+
gr_harassment,
|
1246 |
+
gr_profanity,
|
1247 |
+
gr_political_content,
|
1248 |
+
gr_moral_judgement,
|
1249 |
+
gr_typos
|
1250 |
+
])
|
1251 |
+
gr_score_4_btn.click(lambda x: update_annotation(input_ans_1=x, is_unknown=False, score_value=4),
|
1252 |
+
inputs=[gr_answer_text_2_1],
|
1253 |
+
outputs=[
|
1254 |
+
gr_question_text_1_1,
|
1255 |
+
gr_answer_text_1_1,
|
1256 |
+
gr_question_text_2_1,
|
1257 |
+
gr_answer_text_2_1,
|
1258 |
+
gr_question_text_3_1,
|
1259 |
+
gr_answer_text_3_1,
|
1260 |
+
gr_score_detail,
|
1261 |
+
gr_helphulness,
|
1262 |
+
gr_correcthness,
|
1263 |
+
gr_coherence,
|
1264 |
+
gr_complexity,
|
1265 |
+
gr_verbosity,
|
1266 |
+
gr_humor,
|
1267 |
+
gr_creativity,
|
1268 |
+
gr_appropriate,
|
1269 |
+
gr_following_instructions,
|
1270 |
+
gr_politeness,
|
1271 |
+
gr_harmfulness,
|
1272 |
+
gr_hate,
|
1273 |
+
gr_sexual,
|
1274 |
+
gr_violence,
|
1275 |
+
gr_suicide,
|
1276 |
+
gr_threat,
|
1277 |
+
gr_gun,
|
1278 |
+
gr_controlled_substance,
|
1279 |
+
gr_climinal_planing,
|
1280 |
+
gr_privacy,
|
1281 |
+
gr_harassment,
|
1282 |
+
gr_profanity,
|
1283 |
+
gr_political_content,
|
1284 |
+
gr_moral_judgement,
|
1285 |
+
gr_typos
|
1286 |
+
])
|
1287 |
+
gr_score_5_btn.click(lambda x: update_annotation(input_ans_1=x, is_unknown=False, score_value=5),
|
1288 |
+
inputs=[gr_answer_text_2_1],
|
1289 |
+
outputs=[
|
1290 |
+
gr_question_text_1_1,
|
1291 |
+
gr_answer_text_1_1,
|
1292 |
+
gr_question_text_2_1,
|
1293 |
+
gr_answer_text_2_1,
|
1294 |
+
gr_question_text_3_1,
|
1295 |
+
gr_answer_text_3_1,
|
1296 |
+
gr_score_detail,
|
1297 |
+
gr_helphulness,
|
1298 |
+
gr_correcthness,
|
1299 |
+
gr_coherence,
|
1300 |
+
gr_complexity,
|
1301 |
+
gr_verbosity,
|
1302 |
+
gr_humor,
|
1303 |
+
gr_creativity,
|
1304 |
+
gr_appropriate,
|
1305 |
+
gr_following_instructions,
|
1306 |
+
gr_politeness,
|
1307 |
+
gr_harmfulness,
|
1308 |
+
gr_hate,
|
1309 |
+
gr_sexual,
|
1310 |
+
gr_violence,
|
1311 |
+
gr_suicide,
|
1312 |
+
gr_threat,
|
1313 |
+
gr_gun,
|
1314 |
+
gr_controlled_substance,
|
1315 |
+
gr_climinal_planing,
|
1316 |
+
gr_privacy,
|
1317 |
+
gr_harassment,
|
1318 |
+
gr_profanity,
|
1319 |
+
gr_political_content,
|
1320 |
+
gr_moral_judgement,
|
1321 |
+
gr_typos
|
1322 |
+
])
|
1323 |
+
|
1324 |
+
# Scoreリセット
|
1325 |
+
def score_reset_display():
|
1326 |
+
# gr.State
|
1327 |
+
score.value = 3
|
1328 |
+
helphulness.value = 0
|
1329 |
+
correcthness.value = 0
|
1330 |
+
coherenece.value = 0
|
1331 |
+
complexity.value = 0
|
1332 |
+
verbosity.value = 0
|
1333 |
+
humor.value = 0
|
1334 |
+
creativity.value = 0
|
1335 |
+
appropriate.value = 0
|
1336 |
+
following_instructions.value = 0
|
1337 |
+
politeness.value = 0
|
1338 |
+
harmfulness.value = 0
|
1339 |
+
hate.value = 0
|
1340 |
+
sexual.value = 0
|
1341 |
+
violence.value = 0
|
1342 |
+
suicide.value = 0
|
1343 |
+
threat.value = 0
|
1344 |
+
gun.value = 0
|
1345 |
+
controlled_substance.vlaue = 0
|
1346 |
+
climinal_planing.value = 0
|
1347 |
+
privacy.value = 0
|
1348 |
+
harassment.value = 0
|
1349 |
+
profanity.value = 0
|
1350 |
+
political_content.value = 0
|
1351 |
+
moral_judgement.value = 0
|
1352 |
+
typos.value = 0
|
1353 |
+
|
1354 |
+
return 3,0,0,0,0,0,0,0,0,0,0,0, \
|
1355 |
+
"不明", "不明", "不明", "不明", "不明", \
|
1356 |
+
"不明", "不明", "不明", "不明", "不明", \
|
1357 |
+
"不明", "不明", "不明", "不明"
|
1358 |
+
|
1359 |
+
# Scoreリセット
|
1360 |
+
gr_score_reset.click(
|
1361 |
+
score_reset_display,
|
1362 |
+
inputs=[],
|
1363 |
+
outputs=[
|
1364 |
+
gr_score_detail,
|
1365 |
+
gr_helphulness,
|
1366 |
+
gr_correcthness,
|
1367 |
+
gr_coherence,
|
1368 |
+
gr_complexity,
|
1369 |
+
gr_verbosity,
|
1370 |
+
gr_humor,
|
1371 |
+
gr_creativity,
|
1372 |
+
gr_appropriate,
|
1373 |
+
gr_following_instructions,
|
1374 |
+
gr_politeness,
|
1375 |
+
gr_harmfulness,
|
1376 |
+
gr_hate,
|
1377 |
+
gr_sexual,
|
1378 |
+
gr_violence,
|
1379 |
+
gr_suicide,
|
1380 |
+
gr_threat,
|
1381 |
+
gr_gun,
|
1382 |
+
gr_controlled_substance,
|
1383 |
+
gr_climinal_planing,
|
1384 |
+
gr_privacy,
|
1385 |
+
gr_harassment,
|
1386 |
+
gr_profanity,
|
1387 |
+
gr_political_content,
|
1388 |
+
gr_moral_judgement,
|
1389 |
+
gr_typos
|
1390 |
+
]
|
1391 |
+
)
|
1392 |
+
|
1393 |
+
# データ読込
|
1394 |
+
gr_data_load_btn.click(
|
1395 |
+
dataset_load_fn,
|
1396 |
+
inputs=None,
|
1397 |
+
# textをセットし、intractiveをTrueにする (UI更新)
|
1398 |
+
outputs=[
|
1399 |
+
gr_question_text_1_1,
|
1400 |
+
gr_answer_text_1_1,
|
1401 |
+
gr_question_text_2_1,
|
1402 |
+
gr_answer_text_2_1,
|
1403 |
+
gr_question_text_3_1,
|
1404 |
+
gr_answer_text_3_1,
|
1405 |
+
gr_answer_text_1_1,
|
1406 |
+
gr_answer_text_2_1,
|
1407 |
+
gr_answer_text_3_1,
|
1408 |
+
gr_unknown_btn,
|
1409 |
+
gr_good_btn,
|
1410 |
+
gr_bad_btn,
|
1411 |
+
gr_score_5_btn,
|
1412 |
+
gr_score_4_btn,
|
1413 |
+
gr_score_3_btn,
|
1414 |
+
gr_score_2_btn,
|
1415 |
+
gr_score_1_btn,
|
1416 |
+
gr_submit_score,
|
1417 |
+
gr_score_detail,
|
1418 |
+
gr_score_reset,
|
1419 |
+
gr_helphulness,
|
1420 |
+
gr_correcthness,
|
1421 |
+
gr_coherence,
|
1422 |
+
gr_complexity,
|
1423 |
+
gr_verbosity,
|
1424 |
+
gr_humor,
|
1425 |
+
gr_creativity,
|
1426 |
+
gr_appropriate,
|
1427 |
+
gr_following_instructions,
|
1428 |
+
gr_politeness,
|
1429 |
+
gr_harmfulness,
|
1430 |
+
gr_hate,
|
1431 |
+
gr_sexual,
|
1432 |
+
gr_violence,
|
1433 |
+
gr_suicide,
|
1434 |
+
gr_threat,
|
1435 |
+
gr_gun,
|
1436 |
+
gr_controlled_substance,
|
1437 |
+
gr_climinal_planing,
|
1438 |
+
gr_privacy,
|
1439 |
+
gr_harassment,
|
1440 |
+
gr_profanity,
|
1441 |
+
gr_political_content,
|
1442 |
+
gr_moral_judgement,
|
1443 |
+
gr_typos
|
1444 |
+
]
|
1445 |
+
)
|
1446 |
+
|
1447 |
# TODO Tab切り替えで、アノテ済みの一覧を表示する
|
1448 |
# with gr.Tab("アノテ済みデータセット(管理画面)"):
|
1449 |
# タブを切り替えた時にデータ表示を更新する
|