Spaces:
Sleeping
Sleeping
provide an option to enter ID
Browse files
app.py
CHANGED
@@ -25,7 +25,7 @@ INSTRUCTIONS = """
|
|
25 |
Please compare the two model outputs shown below and select which one you think is better.
|
26 |
- Choose "A is better" if the output from Model A (left) is superior
|
27 |
- Choose "B is better" if the output from Model B (right) is superior
|
28 |
-
- Choose "Tie" if
|
29 |
- Choose "Can't choose" if you cannot make a determination
|
30 |
"""
|
31 |
SAVE_EVERY_N_EXAMPLES = 5
|
@@ -35,11 +35,12 @@ class PairwiseLabeler:
|
|
35 |
def __init__(self):
|
36 |
self.current_index = defaultdict(int)
|
37 |
self.results = defaultdict(list)
|
38 |
-
self.
|
39 |
-
self.
|
|
|
40 |
|
41 |
def __len__(self):
|
42 |
-
return len(self.
|
43 |
|
44 |
def read_hf_dataset(self) -> pd.DataFrame:
|
45 |
try:
|
@@ -70,13 +71,16 @@ class PairwiseLabeler:
|
|
70 |
return pd.DataFrame(sample_data)
|
71 |
|
72 |
def get_current_pair(self, session_id):
|
73 |
-
if
|
|
|
|
|
|
|
74 |
if self.has_url_column:
|
75 |
return None, None, None, None
|
76 |
else:
|
77 |
return None, None, None
|
78 |
|
79 |
-
item = self.df.iloc[self.current_index[session_id]]
|
80 |
item_id = item.get(HF_INPUT_DATASET_ID_COLUMN, f"item_{self.current_index[session_id]}")
|
81 |
left_text = item.get(HF_INPUT_DATASET_COLUMN_A, "")
|
82 |
right_text = item.get(HF_INPUT_DATASET_COLUMN_B, "")
|
@@ -94,11 +98,6 @@ class PairwiseLabeler:
|
|
94 |
else:
|
95 |
return item_id, left_text, right_text, self.current_index[session_id]
|
96 |
|
97 |
-
# Get the current URL if available
|
98 |
-
current_url = None
|
99 |
-
if self.has_url_column:
|
100 |
-
current_url = self.df.iloc[self.current_index[session_id]].get(HF_INPUT_DATASET_URL_COLUMN, "")
|
101 |
-
|
102 |
# Record the judgment
|
103 |
result = {
|
104 |
"item_id": item_id,
|
@@ -106,7 +105,7 @@ class PairwiseLabeler:
|
|
106 |
"timestamp": datetime.datetime.now().isoformat(),
|
107 |
"labeler_id": session_id
|
108 |
}
|
109 |
-
|
110 |
self.results[session_id].append(result)
|
111 |
|
112 |
# Move to next item
|
@@ -159,27 +158,45 @@ class PairwiseLabeler:
|
|
159 |
# Initialize the labeler
|
160 |
labeler = PairwiseLabeler()
|
161 |
|
162 |
-
# Create a
|
163 |
-
def
|
164 |
return str(uuid.uuid4())[:8]
|
165 |
|
166 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
167 |
# State for the session ID
|
168 |
session_id = gr.State(value=None)
|
|
|
169 |
|
170 |
-
#
|
171 |
gr.Markdown(INSTRUCTIONS)
|
172 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
173 |
# URL display component - only shown if URL column is defined
|
174 |
-
|
175 |
if labeler.has_url_column:
|
176 |
-
url_display = gr.HTML(label="Reference URL")
|
177 |
-
|
178 |
-
session_id_display = gr.Textbox(label="Session Information", interactive=False)
|
179 |
|
|
|
180 |
with gr.Row():
|
181 |
with gr.Column():
|
182 |
left_output = gr.Textbox(
|
|
|
183 |
label="Model A Output",
|
184 |
lines=10,
|
185 |
interactive=False
|
@@ -187,68 +204,109 @@ with gr.Blocks() as app:
|
|
187 |
|
188 |
with gr.Column():
|
189 |
right_output = gr.Textbox(
|
|
|
190 |
label="Model B Output",
|
191 |
lines=10,
|
192 |
interactive=False
|
193 |
)
|
194 |
|
195 |
-
item_id = gr.Textbox(visible=False)
|
196 |
|
|
|
197 |
with gr.Row():
|
198 |
-
left_btn = gr.Button("⬅️ A is better", variant="primary")
|
199 |
-
right_btn = gr.Button("➡️ B is better", variant="primary")
|
200 |
-
tie_btn = gr.Button("🤝 Tie", variant="primary")
|
201 |
-
cant_choose_btn = gr.Button("🤔 Can't choose")
|
202 |
-
|
203 |
-
current_sample_sld = gr.Slider(minimum=0, maximum=len(labeler), step=1,
|
204 |
-
interactive=False,
|
205 |
-
label='sample_ind',
|
206 |
-
info=f"Samples labeled (out of {len(labeler)})",
|
207 |
-
show_label=False,
|
208 |
-
container=False,
|
209 |
-
scale=5)
|
210 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
211 |
# Initialize the session and get the first pair
|
212 |
-
def init_session():
|
213 |
-
|
|
|
214 |
|
|
|
215 |
if labeler.has_url_column:
|
216 |
-
initial_id, initial_left, initial_right, initial_url = labeler.get_current_pair(
|
217 |
url_html = f'<a href="{initial_url}" target="_blank">{initial_url}</a>' if initial_url else ""
|
218 |
|
219 |
return (
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
|
|
|
|
|
|
|
|
|
227 |
)
|
228 |
else:
|
229 |
-
initial_id, initial_left, initial_right = labeler.get_current_pair(
|
|
|
|
|
|
|
230 |
|
231 |
return (
|
232 |
-
|
233 |
-
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
|
|
|
|
|
|
|
|
238 |
)
|
239 |
|
240 |
-
#
|
241 |
if labeler.has_url_column:
|
242 |
-
|
243 |
init_session,
|
244 |
-
inputs=
|
245 |
-
outputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
246 |
)
|
247 |
else:
|
248 |
-
|
249 |
init_session,
|
250 |
-
inputs=
|
251 |
-
outputs=[
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
252 |
)
|
253 |
|
254 |
def judge_left(session_id, item_id, left_text, right_text):
|
@@ -264,17 +322,29 @@ with gr.Blocks() as app:
|
|
264 |
return judge("Can't choose", session_id, item_id, left_text, right_text)
|
265 |
|
266 |
def judge(choice, session_id, item_id, left_text, right_text):
|
|
|
|
|
|
|
|
|
267 |
if labeler.has_url_column:
|
268 |
new_id, new_left, new_right, new_url, new_index = labeler.submit_judgment(
|
269 |
item_id, left_text, right_text, choice, session_id
|
270 |
)
|
271 |
url_html = f'<a href="{new_url}" target="_blank">{new_url}</a>' if new_url else ""
|
272 |
-
|
|
|
|
|
|
|
|
|
273 |
else:
|
274 |
new_id, new_left, new_right, new_index = labeler.submit_judgment(
|
275 |
item_id, left_text, right_text, choice, session_id
|
276 |
)
|
277 |
-
|
|
|
|
|
|
|
|
|
278 |
|
279 |
if labeler.has_url_column:
|
280 |
left_btn.click(
|
|
|
25 |
Please compare the two model outputs shown below and select which one you think is better.
|
26 |
- Choose "A is better" if the output from Model A (left) is superior
|
27 |
- Choose "B is better" if the output from Model B (right) is superior
|
28 |
+
- Choose "Tie" if they are equally good or bad
|
29 |
- Choose "Can't choose" if you cannot make a determination
|
30 |
"""
|
31 |
SAVE_EVERY_N_EXAMPLES = 5
|
|
|
35 |
def __init__(self):
|
36 |
self.current_index = defaultdict(int)
|
37 |
self.results = defaultdict(list)
|
38 |
+
self._df = self.read_hf_dataset()
|
39 |
+
self.df = {}
|
40 |
+
self.has_url_column = HF_INPUT_DATASET_URL_COLUMN and HF_INPUT_DATASET_URL_COLUMN in self._df.columns
|
41 |
|
42 |
def __len__(self):
|
43 |
+
return len(self._df)
|
44 |
|
45 |
def read_hf_dataset(self) -> pd.DataFrame:
|
46 |
try:
|
|
|
71 |
return pd.DataFrame(sample_data)
|
72 |
|
73 |
def get_current_pair(self, session_id):
|
74 |
+
if session_id not in self.df:
|
75 |
+
self.df[session_id] = self._df.sample(frac=1).reset_index(drop=True)
|
76 |
+
|
77 |
+
if self.current_index[session_id] >= len(self.df[session_id]):
|
78 |
if self.has_url_column:
|
79 |
return None, None, None, None
|
80 |
else:
|
81 |
return None, None, None
|
82 |
|
83 |
+
item = self.df[session_id].iloc[self.current_index[session_id]]
|
84 |
item_id = item.get(HF_INPUT_DATASET_ID_COLUMN, f"item_{self.current_index[session_id]}")
|
85 |
left_text = item.get(HF_INPUT_DATASET_COLUMN_A, "")
|
86 |
right_text = item.get(HF_INPUT_DATASET_COLUMN_B, "")
|
|
|
98 |
else:
|
99 |
return item_id, left_text, right_text, self.current_index[session_id]
|
100 |
|
|
|
|
|
|
|
|
|
|
|
101 |
# Record the judgment
|
102 |
result = {
|
103 |
"item_id": item_id,
|
|
|
105 |
"timestamp": datetime.datetime.now().isoformat(),
|
106 |
"labeler_id": session_id
|
107 |
}
|
108 |
+
|
109 |
self.results[session_id].append(result)
|
110 |
|
111 |
# Move to next item
|
|
|
158 |
# Initialize the labeler
|
159 |
labeler = PairwiseLabeler()
|
160 |
|
161 |
+
# Create a default session ID
|
162 |
+
def create_default_session():
|
163 |
return str(uuid.uuid4())[:8]
|
164 |
|
165 |
+
# Create a demo session ID for initial display
|
166 |
+
demo_session = create_default_session()
|
167 |
+
demo_item_id, demo_left_text, demo_right_text = None, "", ""
|
168 |
+
demo_url = None
|
169 |
+
if labeler.has_url_column:
|
170 |
+
demo_item_id, demo_left_text, demo_right_text, demo_url = labeler.get_current_pair(demo_session)
|
171 |
+
else:
|
172 |
+
demo_item_id, demo_left_text, demo_right_text = labeler.get_current_pair(demo_session)
|
173 |
+
|
174 |
+
with gr.Blocks(css="footer {visibility: hidden}") as app:
|
175 |
# State for the session ID
|
176 |
session_id = gr.State(value=None)
|
177 |
+
is_session_started = gr.State(value=False)
|
178 |
|
179 |
+
# Header with instructions
|
180 |
gr.Markdown(INSTRUCTIONS)
|
181 |
|
182 |
+
# Session and progress bar in one line
|
183 |
+
user_session_id = gr.Textbox(
|
184 |
+
placeholder="Enter your unique ID or leave blank for random",
|
185 |
+
label="Session ID",
|
186 |
+
scale=2
|
187 |
+
)
|
188 |
+
start_btn = gr.Button("Start Session", variant="primary", scale=1)
|
189 |
+
|
190 |
# URL display component - only shown if URL column is defined
|
191 |
+
print(labeler.has_url_column)
|
192 |
if labeler.has_url_column:
|
193 |
+
url_display = gr.HTML(label="Reference URL", value="", visible=True)
|
|
|
|
|
194 |
|
195 |
+
# Main content area (shown with placeholder text)
|
196 |
with gr.Row():
|
197 |
with gr.Column():
|
198 |
left_output = gr.Textbox(
|
199 |
+
value=demo_left_text,
|
200 |
label="Model A Output",
|
201 |
lines=10,
|
202 |
interactive=False
|
|
|
204 |
|
205 |
with gr.Column():
|
206 |
right_output = gr.Textbox(
|
207 |
+
value=demo_right_text,
|
208 |
label="Model B Output",
|
209 |
lines=10,
|
210 |
interactive=False
|
211 |
)
|
212 |
|
213 |
+
item_id = gr.Textbox(value=demo_item_id, visible=False)
|
214 |
|
215 |
+
# Buttons row (initially disabled)
|
216 |
with gr.Row():
|
217 |
+
left_btn = gr.Button("⬅️ A is better", variant="primary", interactive=False)
|
218 |
+
right_btn = gr.Button("➡️ B is better", variant="primary", interactive=False)
|
219 |
+
tie_btn = gr.Button("🤝 Tie", variant="primary", interactive=False)
|
220 |
+
cant_choose_btn = gr.Button("🤔 Can't choose", interactive=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
221 |
|
222 |
+
# Progress slider (initially disabled)
|
223 |
+
current_sample_sld = gr.Slider(
|
224 |
+
minimum=0,
|
225 |
+
maximum=len(labeler),
|
226 |
+
step=1,
|
227 |
+
value=0,
|
228 |
+
interactive=False,
|
229 |
+
label='Progress',
|
230 |
+
show_label=False
|
231 |
+
)
|
232 |
+
|
233 |
# Initialize the session and get the first pair
|
234 |
+
def init_session(entered_id):
|
235 |
+
# Use entered ID or generate a new one if empty
|
236 |
+
session_id_value = entered_id.strip() if entered_id and entered_id.strip() else create_default_session()
|
237 |
|
238 |
+
# Get the initial data for this session ID
|
239 |
if labeler.has_url_column:
|
240 |
+
initial_id, initial_left, initial_right, initial_url = labeler.get_current_pair(session_id_value)
|
241 |
url_html = f'<a href="{initial_url}" target="_blank">{initial_url}</a>' if initial_url else ""
|
242 |
|
243 |
return (
|
244 |
+
session_id_value, # session_id state
|
245 |
+
True, # is_session_started state
|
246 |
+
labeler.current_index[session_id_value], # current_sample_sld
|
247 |
+
url_html, # url_display
|
248 |
+
initial_left, # left_output
|
249 |
+
initial_right, # right_output
|
250 |
+
initial_id, # item_id
|
251 |
+
gr.update(interactive=True), # left_btn
|
252 |
+
gr.update(interactive=True), # right_btn
|
253 |
+
gr.update(interactive=True), # tie_btn
|
254 |
+
gr.update(interactive=True) # cant_choose_btn
|
255 |
)
|
256 |
else:
|
257 |
+
initial_id, initial_left, initial_right = labeler.get_current_pair(session_id_value)
|
258 |
+
|
259 |
+
# Create session info text
|
260 |
+
progress_text = f"Session: {session_id_value} | {labeler.current_index[session_id_value] + 1}/{len(labeler)}"
|
261 |
|
262 |
return (
|
263 |
+
session_id_value, # session_id state
|
264 |
+
True, # is_session_started state
|
265 |
+
labeler.current_index[session_id_value], # current_sample_sld
|
266 |
+
initial_left, # left_output
|
267 |
+
initial_right, # right_output
|
268 |
+
initial_id, # item_id
|
269 |
+
gr.update(interactive=True), # left_btn
|
270 |
+
gr.update(interactive=True), # right_btn
|
271 |
+
gr.update(interactive=True), # tie_btn
|
272 |
+
gr.update(interactive=True) # cant_choose_btn
|
273 |
)
|
274 |
|
275 |
+
# Connect the start button
|
276 |
if labeler.has_url_column:
|
277 |
+
start_btn.click(
|
278 |
init_session,
|
279 |
+
inputs=[user_session_id],
|
280 |
+
outputs=[
|
281 |
+
session_id,
|
282 |
+
is_session_started,
|
283 |
+
current_sample_sld,
|
284 |
+
url_display,
|
285 |
+
left_output,
|
286 |
+
right_output,
|
287 |
+
item_id,
|
288 |
+
left_btn,
|
289 |
+
right_btn,
|
290 |
+
tie_btn,
|
291 |
+
cant_choose_btn
|
292 |
+
]
|
293 |
)
|
294 |
else:
|
295 |
+
start_btn.click(
|
296 |
init_session,
|
297 |
+
inputs=[user_session_id],
|
298 |
+
outputs=[
|
299 |
+
session_id,
|
300 |
+
is_session_started,
|
301 |
+
current_sample_sld,
|
302 |
+
left_output,
|
303 |
+
right_output,
|
304 |
+
item_id,
|
305 |
+
left_btn,
|
306 |
+
right_btn,
|
307 |
+
tie_btn,
|
308 |
+
cant_choose_btn
|
309 |
+
]
|
310 |
)
|
311 |
|
312 |
def judge_left(session_id, item_id, left_text, right_text):
|
|
|
322 |
return judge("Can't choose", session_id, item_id, left_text, right_text)
|
323 |
|
324 |
def judge(choice, session_id, item_id, left_text, right_text):
|
325 |
+
if not session_id:
|
326 |
+
# Session not initialized, do nothing
|
327 |
+
return item_id, left_text, right_text, 0, "Not started"
|
328 |
+
|
329 |
if labeler.has_url_column:
|
330 |
new_id, new_left, new_right, new_url, new_index = labeler.submit_judgment(
|
331 |
item_id, left_text, right_text, choice, session_id
|
332 |
)
|
333 |
url_html = f'<a href="{new_url}" target="_blank">{new_url}</a>' if new_url else ""
|
334 |
+
|
335 |
+
# Update session info text with new progress
|
336 |
+
progress_text = f"Session: {session_id} | {new_index + 1}/{len(labeler)}" if new_index < len(labeler) else f"Session: {session_id} | Complete! {len(labeler)}/{len(labeler)}"
|
337 |
+
|
338 |
+
return new_id, new_left, new_right, url_html, new_index, progress_text
|
339 |
else:
|
340 |
new_id, new_left, new_right, new_index = labeler.submit_judgment(
|
341 |
item_id, left_text, right_text, choice, session_id
|
342 |
)
|
343 |
+
|
344 |
+
# Update session info text with new progress
|
345 |
+
progress_text = f"Session: {session_id} | {new_index + 1}/{len(labeler)}" if new_index < len(labeler) else f"Session: {session_id} | Complete! {len(labeler)}/{len(labeler)}"
|
346 |
+
|
347 |
+
return new_id, new_left, new_right, new_index, progress_text
|
348 |
|
349 |
if labeler.has_url_column:
|
350 |
left_btn.click(
|