Update app.py
Browse files
app.py
CHANGED
@@ -17,12 +17,12 @@ import time
|
|
17 |
import sys
|
18 |
|
19 |
# Настройка логирования
|
20 |
-
logging.basicConfig(level=logging.INFO)
|
21 |
logger = logging.getLogger("GAIA-Mastermind")
|
22 |
|
23 |
# Конфигурация
|
24 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
25 |
-
MODEL_NAME = "google/flan-t5-large" #
|
26 |
API_RETRIES = 3
|
27 |
API_TIMEOUT = 45
|
28 |
|
@@ -33,12 +33,12 @@ class GAIAThoughtProcessor:
|
|
33 |
logger.info(f"⚡ Инициализация GAIAThoughtProcessor на {self.device.upper()}")
|
34 |
|
35 |
try:
|
36 |
-
# Оптимизированная загрузка модели
|
37 |
self.tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
38 |
self.model = AutoModelForSeq2SeqLM.from_pretrained(
|
39 |
MODEL_NAME,
|
40 |
device_map="auto" if torch.cuda.is_available() else None,
|
41 |
-
torch_dtype=torch.
|
42 |
low_cpu_mem_usage=True
|
43 |
).eval()
|
44 |
|
@@ -47,8 +47,8 @@ class GAIAThoughtProcessor:
|
|
47 |
"text2text-generation",
|
48 |
model=self.model,
|
49 |
tokenizer=self.tokenizer,
|
50 |
-
device
|
51 |
-
max_new_tokens=
|
52 |
)
|
53 |
|
54 |
logger.info("✅ GAIAThoughtProcessor готов")
|
@@ -56,158 +56,30 @@ class GAIAThoughtProcessor:
|
|
56 |
logger.exception("Ошибка инициализации модели")
|
57 |
raise RuntimeError(f"Ошибка инициализации: {str(e)}")
|
58 |
|
59 |
-
def
|
60 |
-
"""
|
61 |
-
try:
|
62 |
-
# Очистка выражения
|
63 |
-
clean_expr = re.sub(r"[^0-9+\-*/().^√π]", "", expression)
|
64 |
-
# Поддержка математических функций
|
65 |
-
context = {
|
66 |
-
"sqrt": np.sqrt,
|
67 |
-
"log": np.log,
|
68 |
-
"log10": np.log10,
|
69 |
-
"pi": np.pi,
|
70 |
-
"e": np.e,
|
71 |
-
"sin": np.sin,
|
72 |
-
"cos": np.cos,
|
73 |
-
"tan": np.tan
|
74 |
-
}
|
75 |
-
return str(eval(clean_expr, {"__builtins__": None}, context))
|
76 |
-
except Exception as e:
|
77 |
-
logger.error(f"Math error: {e}")
|
78 |
-
return f"Math Error: {str(e)}"
|
79 |
-
|
80 |
-
def _table_analyzer(self, table_data: str, query: str) -> str:
|
81 |
-
"""Анализ табличных данных"""
|
82 |
-
try:
|
83 |
-
# Автоопределение формата таблицы
|
84 |
-
if "\t" in table_data:
|
85 |
-
df = pd.read_csv(io.StringIO(table_data), sep="\t")
|
86 |
-
elif "," in table_data:
|
87 |
-
df = pd.read_csv(io.StringIO(table_data))
|
88 |
-
else:
|
89 |
-
df = pd.read_fwf(io.StringIO(table_data))
|
90 |
-
|
91 |
-
# Выполнение запросов
|
92 |
-
query = query.lower()
|
93 |
-
if "sum" in query:
|
94 |
-
return str(df.sum(numeric_only=True).to_dict())
|
95 |
-
elif "mean" in query:
|
96 |
-
return str(df.mean(numeric_only=True).to_dict())
|
97 |
-
elif "max" in query:
|
98 |
-
return str(df.max(numeric_only=True).to_dict())
|
99 |
-
elif "min" in query:
|
100 |
-
return str(df.min(numeric_only=True).to_dict())
|
101 |
-
elif "count" in query:
|
102 |
-
return str(df.count().to_dict())
|
103 |
-
else:
|
104 |
-
return df.describe().to_string()
|
105 |
-
except Exception as e:
|
106 |
-
logger.error(f"Table error: {e}")
|
107 |
-
return f"Table Error: {str(e)}"
|
108 |
-
|
109 |
-
def _text_processor(self, text: str, operation: str) -> str:
|
110 |
-
"""Операции с текстом"""
|
111 |
-
operation = operation.lower()
|
112 |
-
if operation == "reverse":
|
113 |
-
return text[::-1]
|
114 |
-
elif operation == "count_words":
|
115 |
-
return str(len(text.split()))
|
116 |
-
elif operation == "extract_numbers":
|
117 |
-
return ", ".join(re.findall(r"[-+]?\d*\.\d+|\d+", text))
|
118 |
-
elif operation == "uppercase":
|
119 |
-
return text.upper()
|
120 |
-
elif operation == "lowercase":
|
121 |
-
return text.lower()
|
122 |
-
else:
|
123 |
-
return f"Unsupported operation: {operation}"
|
124 |
-
|
125 |
-
def _image_processor(self, image_input: str) -> str:
|
126 |
-
"""Обработка изображений"""
|
127 |
-
try:
|
128 |
-
# Обработка URL
|
129 |
-
if image_input.startswith("http"):
|
130 |
-
response = requests.get(image_input, timeout=30)
|
131 |
-
response.raise_for_status()
|
132 |
-
img_data = response.content
|
133 |
-
img = Image.open(io.BytesIO(img_data))
|
134 |
-
# Обработка base64
|
135 |
-
elif image_input.startswith("data:image"):
|
136 |
-
header, data = image_input.split(",", 1)
|
137 |
-
img_data = base64.b64decode(data)
|
138 |
-
img = Image.open(io.BytesIO(img_data))
|
139 |
-
else:
|
140 |
-
return "Invalid image format"
|
141 |
-
|
142 |
-
# Базовый анализ изображения
|
143 |
-
description = (
|
144 |
-
f"Format: {img.format}, Size: {img.size}, "
|
145 |
-
f"Mode: {img.mode}"
|
146 |
-
)
|
147 |
-
return description
|
148 |
-
except (UnidentifiedImageError, requests.exceptions.RequestException) as e:
|
149 |
-
logger.error(f"Image processing error: {e}")
|
150 |
-
return f"Image Error: {str(e)}"
|
151 |
-
except Exception as e:
|
152 |
-
logger.exception("Unexpected image error")
|
153 |
-
return f"Unexpected Error: {str(e)}"
|
154 |
-
|
155 |
-
def _call_tool(self, tool_name: str, arguments: str) -> str:
|
156 |
-
"""Вызов инструмента по имени"""
|
157 |
try:
|
158 |
-
|
159 |
-
args = [a.strip() for a in arguments.split(",")]
|
160 |
|
161 |
-
if tool_name == "math_solver":
|
162 |
-
return self._math_solver(args[0])
|
163 |
-
elif tool_name == "table_analyzer":
|
164 |
-
return self._table_analyzer(args[0], args[1])
|
165 |
-
elif tool_name == "text_processor":
|
166 |
-
return self._text_processor(args[0], args[1])
|
167 |
-
elif tool_name == "image_processor":
|
168 |
-
return self._image_processor(args[0])
|
169 |
-
else:
|
170 |
-
return f"Unknown tool: {tool_name}"
|
171 |
-
except Exception as e:
|
172 |
-
return f"Tool Error: {str(e)}"
|
173 |
-
|
174 |
-
def _generate_response(self, prompt: str) -> str:
|
175 |
-
"""Генерация ответа с помощью модели"""
|
176 |
-
try:
|
177 |
result = self.text_generator(
|
178 |
prompt,
|
179 |
-
max_new_tokens=
|
180 |
-
num_beams=
|
181 |
early_stopping=True,
|
182 |
-
temperature=0.
|
183 |
)
|
184 |
-
return result[0]['generated_text']
|
185 |
-
except Exception as e:
|
186 |
-
logger.error(f"Generation error: {e}")
|
187 |
-
return f"Generation Error: {str(e)}"
|
188 |
-
finally:
|
189 |
-
# Очистка памяти GPU
|
190 |
-
if "cuda" in self.device:
|
191 |
-
torch.cuda.empty_cache()
|
192 |
-
|
193 |
-
def process_question(self, question: str, task_id: str) -> str:
|
194 |
-
"""Обработка вопроса с декомпозицией на шаги"""
|
195 |
-
try:
|
196 |
-
# Упрощенный промпт для CPU
|
197 |
-
prompt = f"Реши задачу шаг за шагом: {question}\n\nФинальный ответ:"
|
198 |
-
response = self._generate_response(prompt)
|
199 |
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
except Exception as e:
|
206 |
-
logger.
|
207 |
return json.dumps({
|
208 |
"task_id": task_id,
|
209 |
"error": str(e),
|
210 |
-
"final_answer": f"
|
211 |
})
|
212 |
|
213 |
# === СИСТЕМА ОЦЕНКИ ===
|
@@ -224,198 +96,87 @@ class GAIAEvaluationRunner:
|
|
224 |
})
|
225 |
logger.info(f"🌐 Инициализирован GAIAEvaluationRunner для {api_url}")
|
226 |
|
227 |
-
def run_evaluation(self, agent, username: str, agent_code: str, progress=tqdm):
|
228 |
-
# Получение вопросов
|
229 |
-
questions, status = self._fetch_questions()
|
230 |
-
if status != "success":
|
231 |
-
# Возвращаем ошибку в понятном формате
|
232 |
-
error_df = pd.DataFrame([{
|
233 |
-
"Task ID": "ERROR",
|
234 |
-
"Question": status,
|
235 |
-
"Answer": "Не удалось получить вопросы",
|
236 |
-
"Status": "Failed"
|
237 |
-
}])
|
238 |
-
return status, 0, 0, error_df
|
239 |
-
|
240 |
-
# Обработка вопросов
|
241 |
-
results = []
|
242 |
-
answers = []
|
243 |
-
for i, q in enumerate(progress(questions, desc="🧠 Processing GAIA")):
|
244 |
-
try:
|
245 |
-
task_id = q.get("task_id", f"unknown_{i}")
|
246 |
-
json_response = agent.process_question(q["question"], task_id)
|
247 |
-
|
248 |
-
# Парсинг ответа
|
249 |
-
try:
|
250 |
-
response_obj = json.loads(json_response)
|
251 |
-
final_answer = response_obj.get("final_answer", "")
|
252 |
-
if not isinstance(final_answer, str):
|
253 |
-
final_answer = str(final_answer)
|
254 |
-
except json.JSONDecodeError:
|
255 |
-
final_answer = json_response
|
256 |
-
|
257 |
-
# Формирование ответа для GAIA API
|
258 |
-
answers.append({
|
259 |
-
"task_id": task_id,
|
260 |
-
"answer": final_answer[:500] # Ограничение длины
|
261 |
-
})
|
262 |
-
|
263 |
-
# Запись результатов
|
264 |
-
results.append({
|
265 |
-
"Task ID": task_id,
|
266 |
-
"Question": q["question"][:100] + "..." if len(q["question"]) > 100 else q["question"],
|
267 |
-
"Answer": final_answer[:100] + "..." if len(final_answer) > 100 else final_answer,
|
268 |
-
"Status": "Processed"
|
269 |
-
})
|
270 |
-
except Exception as e:
|
271 |
-
logger.error(f"Task {task_id} failed: {e}")
|
272 |
-
answers.append({
|
273 |
-
"task_id": task_id,
|
274 |
-
"answer": f"ERROR: {str(e)}"
|
275 |
-
})
|
276 |
-
results.append({
|
277 |
-
"Task ID": task_id,
|
278 |
-
"Question": "Error",
|
279 |
-
"Answer": f"ERROR: {str(e)}",
|
280 |
-
"Status": "Failed"
|
281 |
-
})
|
282 |
-
|
283 |
-
# Отправка ответов
|
284 |
-
try:
|
285 |
-
submission_result, score = self._submit_answers(username, agent_code, answers)
|
286 |
-
return submission_result, score, len(questions), pd.DataFrame(results)
|
287 |
-
except Exception as e:
|
288 |
-
error_message = f"Ошибка отправки: {str(e)}"
|
289 |
-
results.append({
|
290 |
-
"Task ID": "SUBMIT_ERROR",
|
291 |
-
"Question": error_message,
|
292 |
-
"Answer": "",
|
293 |
-
"Status": "Failed"
|
294 |
-
})
|
295 |
-
return error_message, 0, len(questions), pd.DataFrame(results)
|
296 |
-
|
297 |
def _fetch_questions(self) -> Tuple[list, str]:
|
298 |
"""Получение вопросов с API"""
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
-
|
306 |
-
|
307 |
-
|
308 |
-
|
309 |
-
|
310 |
-
|
311 |
-
|
312 |
-
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
elif response.status_code == 429:
|
317 |
-
wait_time = 2 ** attempt # Экспоненциальная задержка
|
318 |
-
logger.warning(f"Rate limited, retrying in {wait_time}s...")
|
319 |
-
time.sleep(wait_time)
|
320 |
-
continue
|
321 |
-
|
322 |
-
else:
|
323 |
-
return [], f"Ошибка API: HTTP {response.status_code} - {response.text}"
|
324 |
|
325 |
-
|
326 |
-
|
327 |
-
|
328 |
-
|
329 |
-
logger.error(f"Неожиданная ошибка: {e}")
|
330 |
-
return [], f"Неожиданная ошибка: {str(e)}"
|
331 |
-
|
332 |
-
return [], "API недоступен после попыток"
|
333 |
|
334 |
def _submit_answers(self, username: str, agent_code: str, answers: list) -> Tuple[str, int]:
|
335 |
"""Отправка ответов на сервер"""
|
336 |
-
|
337 |
-
|
338 |
-
|
339 |
-
|
340 |
-
|
341 |
-
|
342 |
-
|
343 |
-
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
|
348 |
-
|
349 |
-
|
350 |
-
|
351 |
-
|
352 |
-
|
353 |
-
|
354 |
-
|
355 |
-
|
356 |
-
|
357 |
-
|
358 |
-
|
359 |
-
|
360 |
-
|
361 |
-
|
362 |
-
logger.warning(f"Rate limited, retrying in {wait_time}s...")
|
363 |
-
time.sleep(wait_time)
|
364 |
-
continue
|
365 |
-
|
366 |
-
else:
|
367 |
-
return f"HTTP Ошибка {response.status_code} - {response.text}", 0
|
368 |
|
369 |
-
|
370 |
-
|
371 |
-
|
372 |
-
|
373 |
-
logger.error(f"Неожиданная ошибка отправки: {e}")
|
374 |
-
return f"Неожиданная ошибка: {str(e)}", 0
|
375 |
-
|
376 |
-
return "Сбой отправки после попыток", 0
|
377 |
|
378 |
-
|
379 |
-
|
380 |
-
try:
|
381 |
-
progress(0, desc="⚡ Инициализация GAIA Mastermind...")
|
382 |
-
agent = GAIAThoughtProcessor()
|
383 |
-
|
384 |
-
progress(0.1, desc="🌐 Подключение к GAIA API...")
|
385 |
-
runner = GAIAEvaluationRunner()
|
386 |
-
|
387 |
# Получение вопросов
|
388 |
-
progress(0.
|
389 |
-
questions, status =
|
390 |
if status != "success":
|
391 |
-
|
392 |
-
error_df = pd.DataFrame([{
|
393 |
-
"Task ID": "ERROR",
|
394 |
-
"Question": error_message,
|
395 |
-
"Answer": "Не удалось получить вопросы",
|
396 |
-
"Status": "Failed"
|
397 |
-
}])
|
398 |
-
return error_message, 0, 0, error_df
|
399 |
|
400 |
-
|
401 |
-
if
|
402 |
-
|
403 |
-
error_df = pd.DataFrame([{
|
404 |
-
"Task ID": "ERROR",
|
405 |
-
"Question": error_message,
|
406 |
-
"Answer": "Нет данных",
|
407 |
-
"Status": "Failed"
|
408 |
-
}])
|
409 |
-
return error_message, 0, 0, error_df
|
410 |
|
411 |
-
# Обработка вопросов
|
412 |
results = []
|
413 |
answers = []
|
414 |
|
415 |
for i, q in enumerate(questions):
|
416 |
-
progress(i /
|
417 |
try:
|
418 |
-
task_id = q.get("task_id", f"
|
|
|
|
|
419 |
json_response = agent.process_question(q["question"], task_id)
|
420 |
|
421 |
# Парсинг ответа
|
@@ -432,12 +193,12 @@ def run_evaluation(username: str, agent_code: str, progress=gr.Progress()):
|
|
432 |
|
433 |
results.append({
|
434 |
"Task ID": task_id,
|
435 |
-
"Question": q["question"][:
|
436 |
-
"Answer": str(final_answer)[:
|
437 |
"Status": "Processed"
|
438 |
})
|
439 |
except Exception as e:
|
440 |
-
logger.error(f"
|
441 |
answers.append({
|
442 |
"task_id": task_id,
|
443 |
"answer": f"ERROR: {str(e)}"
|
@@ -450,84 +211,63 @@ def run_evaluation(username: str, agent_code: str, progress=gr.Progress()):
|
|
450 |
})
|
451 |
|
452 |
# Отправка ответов
|
453 |
-
progress(0.9, desc="
|
454 |
-
submission_result, score =
|
455 |
-
return submission_result, score,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
456 |
|
457 |
except Exception as e:
|
458 |
-
logger.exception("
|
459 |
-
error_message = f"Критическая ошибка: {str(e)}"
|
460 |
error_df = pd.DataFrame([{
|
461 |
-
"Task ID": "
|
462 |
-
"Question":
|
463 |
"Answer": "См. логи",
|
464 |
"Status": "Failed"
|
465 |
}])
|
466 |
-
return
|
467 |
|
468 |
# Создание интерфейса
|
469 |
-
with gr.Blocks(
|
470 |
-
|
471 |
-
|
472 |
-
css="""
|
473 |
-
.gradio-container {background: linear-gradient(135deg, #1a2a6c, #2c5364)}
|
474 |
-
.dark {color: #f0f0f0}
|
475 |
-
"""
|
476 |
-
) as demo:
|
477 |
-
gr.Markdown("""
|
478 |
-
<div style="text-align:center; background: linear-gradient(135deg, #0f2027, #203a43);
|
479 |
-
padding: 20px; border-radius: 15px; color: white; box-shadow: 0 10px 20px rgba(0,0,0,0.3);">
|
480 |
-
<h1>🧠 GAIA Mastermind</h1>
|
481 |
-
<h3>Многошаговое решение задач с декомпозицией</h3>
|
482 |
-
<p>Соответствует спецификации GAIA API</p>
|
483 |
-
</div>
|
484 |
-
""")
|
485 |
|
486 |
with gr.Row():
|
487 |
-
with gr.Column(
|
488 |
-
gr.Markdown("
|
489 |
-
username = gr.Textbox(
|
490 |
-
|
491 |
-
|
492 |
-
info="Ваше имя пользователя Hugging Face"
|
493 |
-
)
|
494 |
-
agent_code = gr.Textbox(
|
495 |
-
label="Agent Code",
|
496 |
-
value="https://huggingface.co/spaces/yoshizen/FinalTest",
|
497 |
-
info="URL вашего агента"
|
498 |
-
)
|
499 |
-
run_btn = gr.Button("🚀 Запустить оценку", variant="primary", scale=1)
|
500 |
|
501 |
-
gr.Markdown("
|
502 |
-
sys_info = gr.Textbox(label="Системная информация", interactive=False
|
503 |
|
504 |
-
with gr.Column(
|
505 |
-
gr.Markdown("
|
506 |
with gr.Row():
|
507 |
-
result_output = gr.Textbox(
|
508 |
-
|
509 |
-
|
510 |
-
max_lines=3
|
511 |
-
)
|
512 |
-
correct_output = gr.Number(
|
513 |
-
label="✅ Правильные ответы",
|
514 |
-
interactive=False
|
515 |
-
)
|
516 |
-
total_output = gr.Number(
|
517 |
-
label="📚 Всего вопросов",
|
518 |
-
interactive=False
|
519 |
-
)
|
520 |
|
521 |
-
# Упрощенный Dataframe
|
522 |
results_table = gr.Dataframe(
|
523 |
-
label="
|
524 |
headers=["Task ID", "Question", "Answer", "Status"],
|
525 |
interactive=False
|
526 |
)
|
527 |
|
528 |
# Системная информация
|
529 |
def get_system_info():
|
530 |
-
device = "GPU
|
531 |
return f"Device: {device} | Model: {MODEL_NAME} | API: {DEFAULT_API_URL}"
|
532 |
|
533 |
demo.load(get_system_info, inputs=None, outputs=sys_info)
|
@@ -536,15 +276,13 @@ with gr.Blocks(
|
|
536 |
fn=run_evaluation,
|
537 |
inputs=[username, agent_code],
|
538 |
outputs=[result_output, correct_output, total_output, results_table],
|
539 |
-
concurrency_limit=1
|
540 |
-
show_progress="minimal"
|
541 |
)
|
542 |
|
543 |
if __name__ == "__main__":
|
544 |
-
demo.queue(max_size=
|
545 |
server_name="0.0.0.0",
|
546 |
server_port=7860,
|
547 |
share=False,
|
548 |
-
show_error=True
|
549 |
-
debug=True # Включение детального лога
|
550 |
)
|
|
|
17 |
import sys
|
18 |
|
19 |
# Настройка логирования
|
20 |
+
logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
|
21 |
logger = logging.getLogger("GAIA-Mastermind")
|
22 |
|
23 |
# Конфигурация
|
24 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
25 |
+
MODEL_NAME = "google/flan-t5-large" # Оптимизировано для CPU
|
26 |
API_RETRIES = 3
|
27 |
API_TIMEOUT = 45
|
28 |
|
|
|
33 |
logger.info(f"⚡ Инициализация GAIAThoughtProcessor на {self.device.upper()}")
|
34 |
|
35 |
try:
|
36 |
+
# Оптимизированная загрузка модели для CPU
|
37 |
self.tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
38 |
self.model = AutoModelForSeq2SeqLM.from_pretrained(
|
39 |
MODEL_NAME,
|
40 |
device_map="auto" if torch.cuda.is_available() else None,
|
41 |
+
torch_dtype=torch.float32,
|
42 |
low_cpu_mem_usage=True
|
43 |
).eval()
|
44 |
|
|
|
47 |
"text2text-generation",
|
48 |
model=self.model,
|
49 |
tokenizer=self.tokenizer,
|
50 |
+
device=-1 if self.device == "cpu" else 0,
|
51 |
+
max_new_tokens=128
|
52 |
)
|
53 |
|
54 |
logger.info("✅ GAIAThoughtProcessor готов")
|
|
|
56 |
logger.exception("Ошибка инициализации модели")
|
57 |
raise RuntimeError(f"Ошибка инициализации: {str(e)}")
|
58 |
|
59 |
+
def process_question(self, question: str, task_id: str) -> str:
|
60 |
+
"""Упрощенная обработка вопроса"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
61 |
try:
|
62 |
+
prompt = f"Реши задачу шаг за шагом: {question}\n\nФинальный ответ:"
|
|
|
63 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
64 |
result = self.text_generator(
|
65 |
prompt,
|
66 |
+
max_new_tokens=128,
|
67 |
+
num_beams=2,
|
68 |
early_stopping=True,
|
69 |
+
temperature=0.1
|
70 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
71 |
|
72 |
+
response = result[0]['generated_text'].strip()
|
73 |
+
|
74 |
+
# Создаем JSON ответ
|
75 |
+
return json.dumps({"final_answer": response})
|
76 |
+
|
77 |
except Exception as e:
|
78 |
+
logger.error(f"Ошибка обработки вопроса: {str(e)}")
|
79 |
return json.dumps({
|
80 |
"task_id": task_id,
|
81 |
"error": str(e),
|
82 |
+
"final_answer": f"ERROR: {str(e)}"
|
83 |
})
|
84 |
|
85 |
# === СИСТЕМА ОЦЕНКИ ===
|
|
|
96 |
})
|
97 |
logger.info(f"🌐 Инициализирован GAIAEvaluationRunner для {api_url}")
|
98 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
99 |
def _fetch_questions(self) -> Tuple[list, str]:
|
100 |
"""Получение вопросов с API"""
|
101 |
+
logger.info(f"🔍 Запрос вопросов с {self.questions_url}")
|
102 |
+
try:
|
103 |
+
response = self.session.get(
|
104 |
+
self.questions_url,
|
105 |
+
timeout=API_TIMEOUT
|
106 |
+
)
|
107 |
+
|
108 |
+
logger.info(f"Статус ответа: {response.status_code}")
|
109 |
+
|
110 |
+
if response.status_code == 200:
|
111 |
+
questions = response.json()
|
112 |
+
logger.info(f"Получено {len(questions)} вопросов")
|
113 |
+
return questions, "success"
|
114 |
+
else:
|
115 |
+
error_msg = f"Ошибка API: HTTP {response.status_code}"
|
116 |
+
logger.error(error_msg)
|
117 |
+
return [], error_msg
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
118 |
|
119 |
+
except Exception as e:
|
120 |
+
error_msg = f"Ошибка соединения: {str(e)}"
|
121 |
+
logger.exception(error_msg)
|
122 |
+
return [], error_msg
|
|
|
|
|
|
|
|
|
123 |
|
124 |
def _submit_answers(self, username: str, agent_code: str, answers: list) -> Tuple[str, int]:
|
125 |
"""Отправка ответов на сервер"""
|
126 |
+
logger.info(f"📤 Отправка ответов для пользователя {username}")
|
127 |
+
try:
|
128 |
+
payload = {
|
129 |
+
"username": username.strip(),
|
130 |
+
"agent_code": agent_code.strip(),
|
131 |
+
"answers": answers
|
132 |
+
}
|
133 |
+
|
134 |
+
response = self.session.post(
|
135 |
+
self.submit_url,
|
136 |
+
json=payload,
|
137 |
+
timeout=API_TIMEOUT * 2
|
138 |
+
)
|
139 |
+
|
140 |
+
logger.info(f"Статус отправки: {response.status_code}")
|
141 |
+
|
142 |
+
if response.status_code == 200:
|
143 |
+
result = response.json()
|
144 |
+
score = result.get("score", 0)
|
145 |
+
return result.get("message", "Ответы успешно отправлены"), score
|
146 |
+
else:
|
147 |
+
error = f"HTTP Ошибка {response.status_code}"
|
148 |
+
if response.text:
|
149 |
+
error += f": {response.text[:200]}"
|
150 |
+
logger.error(error)
|
151 |
+
return error, 0
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
|
153 |
+
except Exception as e:
|
154 |
+
error = f"Ошибка отправки: {str(e)}"
|
155 |
+
logger.exception(error)
|
156 |
+
return error, 0
|
|
|
|
|
|
|
|
|
157 |
|
158 |
+
def run_evaluation(self, agent, username: str, agent_code: str, progress=gr.Progress()):
|
159 |
+
"""Основной процесс оценки"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
160 |
# Получение вопросов
|
161 |
+
progress(0.1, desc="Получение вопросов")
|
162 |
+
questions, status = self._fetch_questions()
|
163 |
if status != "success":
|
164 |
+
return status, 0, 0, pd.DataFrame()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
165 |
|
166 |
+
total_questions = len(questions)
|
167 |
+
if total_questions == 0:
|
168 |
+
return "Получено 0 вопросов", 0, 0, pd.DataFrame()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
169 |
|
170 |
+
# Обработка вопросов
|
171 |
results = []
|
172 |
answers = []
|
173 |
|
174 |
for i, q in enumerate(questions):
|
175 |
+
progress(i / total_questions, desc=f"Обработка задачи {i+1}/{total_questions}")
|
176 |
try:
|
177 |
+
task_id = q.get("task_id", f"task_{i}")
|
178 |
+
logger.info(f"🔧 Обработка задачи {task_id}")
|
179 |
+
|
180 |
json_response = agent.process_question(q["question"], task_id)
|
181 |
|
182 |
# Парсинг ответа
|
|
|
193 |
|
194 |
results.append({
|
195 |
"Task ID": task_id,
|
196 |
+
"Question": q["question"][:50] + "..." if len(q["question"]) > 50 else q["question"],
|
197 |
+
"Answer": str(final_answer)[:50] + "..." if len(str(final_answer)) > 50 else str(final_answer),
|
198 |
"Status": "Processed"
|
199 |
})
|
200 |
except Exception as e:
|
201 |
+
logger.error(f"Ошибка обработки задачи: {str(e)}")
|
202 |
answers.append({
|
203 |
"task_id": task_id,
|
204 |
"answer": f"ERROR: {str(e)}"
|
|
|
211 |
})
|
212 |
|
213 |
# Отправка ответов
|
214 |
+
progress(0.9, desc="Отправка результатов")
|
215 |
+
submission_result, score = self._submit_answers(username, agent_code, answers)
|
216 |
+
return submission_result, score, total_questions, pd.DataFrame(results)
|
217 |
+
|
218 |
+
# === ИНТЕРФЕЙС GRADIO ===
|
219 |
+
def run_evaluation(username: str, agent_code: str, progress=gr.Progress()):
|
220 |
+
try:
|
221 |
+
progress(0, desc="Инициализация агента")
|
222 |
+
agent = GAIAThoughtProcessor()
|
223 |
+
|
224 |
+
progress(0.1, desc="Подключение к API")
|
225 |
+
runner = GAIAEvaluationRunner()
|
226 |
+
|
227 |
+
# Запуск оценки
|
228 |
+
return runner.run_evaluation(agent, username, agent_code, progress)
|
229 |
|
230 |
except Exception as e:
|
231 |
+
logger.exception("Критическая ошибка в run_evaluation")
|
|
|
232 |
error_df = pd.DataFrame([{
|
233 |
+
"Task ID": "ERROR",
|
234 |
+
"Question": f"Критическая ошибка: {str(e)}",
|
235 |
"Answer": "См. логи",
|
236 |
"Status": "Failed"
|
237 |
}])
|
238 |
+
return f"Ошибка: {str(e)}", 0, 0, error_df
|
239 |
|
240 |
# Создание интерфейса
|
241 |
+
with gr.Blocks(title="GAIA Mastermind") as demo:
|
242 |
+
gr.Markdown("# GAIA Mastermind")
|
243 |
+
gr.Markdown("Многошаговое решение задач с декомпозицией")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
244 |
|
245 |
with gr.Row():
|
246 |
+
with gr.Column():
|
247 |
+
gr.Markdown("## 🔐 Авторизация")
|
248 |
+
username = gr.Textbox(label="HF Username", value="yoshizen")
|
249 |
+
agent_code = gr.Textbox(label="Agent Code", value="https://huggingface.co/spaces/yoshizen/FinalTest")
|
250 |
+
run_btn = gr.Button("Запустить оценку")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
251 |
|
252 |
+
gr.Markdown("## ⚙️ Статус системы")
|
253 |
+
sys_info = gr.Textbox(label="Системная информация", interactive=False)
|
254 |
|
255 |
+
with gr.Column():
|
256 |
+
gr.Markdown("## 📊 Результаты GAIA")
|
257 |
with gr.Row():
|
258 |
+
result_output = gr.Textbox(label="Статус отправки", interactive=False)
|
259 |
+
correct_output = gr.Number(label="Правильные ответы", interactive=False)
|
260 |
+
total_output = gr.Number(label="Всего вопросов", interactive=False)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
261 |
|
|
|
262 |
results_table = gr.Dataframe(
|
263 |
+
label="Детализация ответов",
|
264 |
headers=["Task ID", "Question", "Answer", "Status"],
|
265 |
interactive=False
|
266 |
)
|
267 |
|
268 |
# Системная информация
|
269 |
def get_system_info():
|
270 |
+
device = "GPU" if torch.cuda.is_available() else "CPU"
|
271 |
return f"Device: {device} | Model: {MODEL_NAME} | API: {DEFAULT_API_URL}"
|
272 |
|
273 |
demo.load(get_system_info, inputs=None, outputs=sys_info)
|
|
|
276 |
fn=run_evaluation,
|
277 |
inputs=[username, agent_code],
|
278 |
outputs=[result_output, correct_output, total_output, results_table],
|
279 |
+
concurrency_limit=1
|
|
|
280 |
)
|
281 |
|
282 |
if __name__ == "__main__":
|
283 |
+
demo.queue(max_size=1).launch(
|
284 |
server_name="0.0.0.0",
|
285 |
server_port=7860,
|
286 |
share=False,
|
287 |
+
show_error=True
|
|
|
288 |
)
|