Update app.py
Browse files
app.py
CHANGED
|
@@ -7,11 +7,28 @@ import gradio as gr
|
|
| 7 |
import requests
|
| 8 |
import pandas as pd
|
| 9 |
import spaces
|
|
|
|
| 10 |
from smolagents import CodeAgent, DuckDuckGoSearchTool, VisitWebpageTool, OpenAIServerModel, tool
|
| 11 |
from huggingface_hub import InferenceClient
|
| 12 |
|
| 13 |
# --- Constants ---
|
| 14 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
# Satisfies ZeroGPU scanner requirement
|
| 17 |
@spaces.GPU
|
|
@@ -53,6 +70,25 @@ def transcribe_audio(file_path: str) -> str:
|
|
| 53 |
Args:
|
| 54 |
file_path: Local path to the audio file.
|
| 55 |
"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 56 |
token = os.getenv("HF_TOKEN")
|
| 57 |
client = InferenceClient(token=token)
|
| 58 |
try:
|
|
@@ -70,28 +106,114 @@ def transcribe_audio(file_path: str) -> str:
|
|
| 70 |
except Exception as e:
|
| 71 |
return f"Speech-to-text transcription failed: {e}"
|
| 72 |
|
| 73 |
-
# --- Model &
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 74 |
|
| 75 |
def build_gaia_agent() -> CodeAgent:
|
| 76 |
-
"""Build CodeAgent using Groq
|
| 77 |
groq_key = os.getenv("GROQ_API_KEY")
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
print("Using Groq API provider (llama-3.3-70b-versatile)...")
|
| 81 |
-
model = OpenAIServerModel(
|
| 82 |
-
model_id="llama-3.3-70b-versatile",
|
| 83 |
-
api_base="https://api.groq.com/openai/v1",
|
| 84 |
-
api_key=groq_key
|
| 85 |
-
)
|
| 86 |
-
else:
|
| 87 |
-
print("Warning: GROQ_API_KEY not found. Falling back to HF Inference...")
|
| 88 |
-
token = os.getenv("HF_TOKEN")
|
| 89 |
-
model = OpenAIServerModel(
|
| 90 |
-
model_id="Qwen/Qwen2.5-Coder-32B-Instruct",
|
| 91 |
-
api_base="https://router.huggingface.co/v1",
|
| 92 |
-
api_key=token
|
| 93 |
-
)
|
| 94 |
|
|
|
|
| 95 |
search_tool = DuckDuckGoSearchTool()
|
| 96 |
web_visit_tool = VisitWebpageTool()
|
| 97 |
|
|
@@ -124,12 +246,12 @@ def build_gaia_agent() -> CodeAgent:
|
|
| 124 |
"zipfile",
|
| 125 |
"base64"
|
| 126 |
],
|
| 127 |
-
max_steps=
|
| 128 |
)
|
| 129 |
return agent
|
| 130 |
|
| 131 |
class GAIAgentWrapper:
|
| 132 |
-
"""Wrapper to run agent
|
| 133 |
def __init__(self):
|
| 134 |
print("Initializing CodeAgent...")
|
| 135 |
self.agent = build_gaia_agent()
|
|
@@ -147,39 +269,29 @@ class GAIAgentWrapper:
|
|
| 147 |
"6. Never start with 'FINAL ANSWER:'. Output ONLY the pure value."
|
| 148 |
)
|
| 149 |
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
cleaned = str(raw_result).strip()
|
| 154 |
-
|
| 155 |
-
# Clean common prefixes
|
| 156 |
-
for prefix in ["final answer:", "answer:", "the answer is"]:
|
| 157 |
-
if cleaned.lower().startswith(prefix):
|
| 158 |
-
cleaned = cleaned[len(prefix):].strip(" :\n*`")
|
| 159 |
-
|
| 160 |
-
cleaned = cleaned.strip(" '\"`*")
|
| 161 |
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
cleaned = str(int(as_float))
|
| 167 |
-
except ValueError:
|
| 168 |
-
pass
|
| 169 |
|
| 170 |
-
|
| 171 |
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
print(f"Agent execution error: {e}")
|
| 180 |
-
return f"AGENT ERROR: {e}"
|
| 181 |
|
| 182 |
-
|
|
|
|
|
|
|
|
|
|
| 183 |
|
| 184 |
# --- File Handling ---
|
| 185 |
|
|
@@ -278,13 +390,14 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
|
|
| 278 |
results_log = []
|
| 279 |
answers_payload = []
|
| 280 |
print(f"Running agent on {len(questions_data)} questions...")
|
| 281 |
-
for item in questions_data:
|
| 282 |
task_id = item.get("task_id")
|
| 283 |
question_text = item.get("question")
|
| 284 |
suggested_filename = item.get("file_name")
|
| 285 |
if not task_id or question_text is None:
|
| 286 |
continue
|
| 287 |
|
|
|
|
| 288 |
file_path = download_task_file(api_url, task_id, suggested_name=suggested_filename)
|
| 289 |
if file_path:
|
| 290 |
file_ext = os.path.splitext(file_path)[-1]
|
|
@@ -304,7 +417,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
|
|
| 304 |
print(f"Error running agent on task {task_id}: {e}")
|
| 305 |
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"AGENT ERROR: {e}"})
|
| 306 |
|
| 307 |
-
time.sleep(
|
| 308 |
|
| 309 |
if not answers_payload:
|
| 310 |
return "Agent did not produce any answers to submit.", pd.DataFrame(results_log)
|
|
|
|
| 7 |
import requests
|
| 8 |
import pandas as pd
|
| 9 |
import spaces
|
| 10 |
+
from openai import OpenAI
|
| 11 |
from smolagents import CodeAgent, DuckDuckGoSearchTool, VisitWebpageTool, OpenAIServerModel, tool
|
| 12 |
from huggingface_hub import InferenceClient
|
| 13 |
|
| 14 |
# --- Constants ---
|
| 15 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
| 16 |
+
GROQ_BASE_URL = "https://api.groq.com/openai/v1"
|
| 17 |
+
|
| 18 |
+
# Preferred order of active Groq reasoning/coding models
|
| 19 |
+
PREFERRED_GROQ_MODELS = [
|
| 20 |
+
"openai/gpt-oss-120b",
|
| 21 |
+
"openai/gpt-oss-20b",
|
| 22 |
+
"qwen/qwen3.8-27b",
|
| 23 |
+
"qwen/qwen3.6-27b",
|
| 24 |
+
"qwen/qwen3-32b",
|
| 25 |
+
"qwen-qwq-32b",
|
| 26 |
+
"deepseek-r1-distill-llama-70b",
|
| 27 |
+
"meta-llama/llama-4-maverick-17b-128e-instruct",
|
| 28 |
+
"meta-llama/llama-4-scout-17b-16e-instruct",
|
| 29 |
+
"llama-3.3-70b-versatile",
|
| 30 |
+
"llama-3.1-8b-instant",
|
| 31 |
+
]
|
| 32 |
|
| 33 |
# Satisfies ZeroGPU scanner requirement
|
| 34 |
@spaces.GPU
|
|
|
|
| 70 |
Args:
|
| 71 |
file_path: Local path to the audio file.
|
| 72 |
"""
|
| 73 |
+
groq_key = os.getenv("GROQ_API_KEY")
|
| 74 |
+
# Priority 1: Use Groq's free Whisper endpoint
|
| 75 |
+
if groq_key:
|
| 76 |
+
try:
|
| 77 |
+
client = OpenAI(base_url=GROQ_BASE_URL, api_key=groq_key)
|
| 78 |
+
for whisper_model in ["whisper-large-v3-turbo", "whisper-large-v3"]:
|
| 79 |
+
try:
|
| 80 |
+
with open(file_path, "rb") as audio_file:
|
| 81 |
+
transcription = client.audio.transcriptions.create(
|
| 82 |
+
model=whisper_model,
|
| 83 |
+
file=audio_file
|
| 84 |
+
)
|
| 85 |
+
return transcription.text
|
| 86 |
+
except Exception as inner_e:
|
| 87 |
+
print(f"Groq Whisper ({whisper_model}) error: {inner_e}")
|
| 88 |
+
except Exception as e:
|
| 89 |
+
print(f"Groq audio transcription error: {e}")
|
| 90 |
+
|
| 91 |
+
# Priority 2: Fallback to Hugging Face InferenceClient
|
| 92 |
token = os.getenv("HF_TOKEN")
|
| 93 |
client = InferenceClient(token=token)
|
| 94 |
try:
|
|
|
|
| 106 |
except Exception as e:
|
| 107 |
return f"Speech-to-text transcription failed: {e}"
|
| 108 |
|
| 109 |
+
# --- Dynamic Model Discovery & Per-Step Resilient Wrapper ---
|
| 110 |
+
|
| 111 |
+
def discover_groq_models(api_key: str) -> list[str]:
|
| 112 |
+
"""Query Groq API for active models on this key and verify with a live ping."""
|
| 113 |
+
client = OpenAI(base_url=GROQ_BASE_URL, api_key=api_key)
|
| 114 |
+
available_ids = []
|
| 115 |
+
try:
|
| 116 |
+
models_resp = client.models.list()
|
| 117 |
+
available_ids = [m.id for m in models_resp.data]
|
| 118 |
+
print(f"Groq models available on account: {available_ids}")
|
| 119 |
+
except Exception as e:
|
| 120 |
+
print(f"Could not list Groq models, using fallback list: {e}")
|
| 121 |
+
available_ids = PREFERRED_GROQ_MODELS
|
| 122 |
+
|
| 123 |
+
# Filter out non-chat models (whisper, guard, tts, embed)
|
| 124 |
+
excluded_keywords = ["whisper", "guard", "safeguard", "tts", "orpheus", "embed", "playai"]
|
| 125 |
+
chat_candidates = []
|
| 126 |
+
|
| 127 |
+
# First add preferred models that exist in available_ids
|
| 128 |
+
for pref in PREFERRED_GROQ_MODELS:
|
| 129 |
+
if pref in available_ids and pref not in chat_candidates:
|
| 130 |
+
chat_candidates.append(pref)
|
| 131 |
+
|
| 132 |
+
# Then add any remaining chat models reported by the API
|
| 133 |
+
for mid in available_ids:
|
| 134 |
+
if mid not in chat_candidates and not any(k in mid.lower() for k in excluded_keywords):
|
| 135 |
+
chat_candidates.append(mid)
|
| 136 |
+
|
| 137 |
+
# Verify up to 3 working models with a 1-token test call
|
| 138 |
+
verified = []
|
| 139 |
+
for candidate in chat_candidates:
|
| 140 |
+
try:
|
| 141 |
+
client.chat.completions.create(
|
| 142 |
+
model=candidate,
|
| 143 |
+
messages=[{"role": "user", "content": "Reply with 1"}],
|
| 144 |
+
max_tokens=16
|
| 145 |
+
)
|
| 146 |
+
print(f"Verified working Groq model: {candidate}")
|
| 147 |
+
verified.append(candidate)
|
| 148 |
+
if len(verified) >= 3:
|
| 149 |
+
break
|
| 150 |
+
except Exception as e:
|
| 151 |
+
print(f"Skipping model {candidate} (ping failed): {e}")
|
| 152 |
+
|
| 153 |
+
return verified
|
| 154 |
+
|
| 155 |
+
def create_resilient_groq_model(api_key: str) -> OpenAIServerModel:
|
| 156 |
+
"""Create an OpenAIServerModel with per-step retry and automatic model failover."""
|
| 157 |
+
verified_models = discover_groq_models(api_key)
|
| 158 |
+
if not verified_models:
|
| 159 |
+
raise RuntimeError("No active Groq chat models responded to ping test. Check GROQ_API_KEY.")
|
| 160 |
+
|
| 161 |
+
primary_model = verified_models[0]
|
| 162 |
+
print(f"Primary model set to: {primary_model} | Backups: {verified_models[1:]}")
|
| 163 |
+
|
| 164 |
+
model = OpenAIServerModel(
|
| 165 |
+
model_id=primary_model,
|
| 166 |
+
api_base=GROQ_BASE_URL,
|
| 167 |
+
api_key=api_key
|
| 168 |
+
)
|
| 169 |
+
|
| 170 |
+
original_create = model.client.chat.completions.create
|
| 171 |
+
|
| 172 |
+
def resilient_create(*args, **kwargs):
|
| 173 |
+
# Gentle pacing between LLM steps to avoid hitting 30 RPM / 8000 TPM limits
|
| 174 |
+
time.sleep(2.0)
|
| 175 |
+
|
| 176 |
+
last_err = None
|
| 177 |
+
for attempt in range(5):
|
| 178 |
+
# Rotate to backup models if primary hits rate limit
|
| 179 |
+
active_model = verified_models[attempt % len(verified_models)]
|
| 180 |
+
kwargs["model"] = active_model
|
| 181 |
+
|
| 182 |
+
try:
|
| 183 |
+
return original_create(*args, **kwargs)
|
| 184 |
+
except Exception as e:
|
| 185 |
+
last_err = e
|
| 186 |
+
err_str = str(e).lower()
|
| 187 |
+
|
| 188 |
+
# Handle unsupported 'stop' parameter on reasoning models
|
| 189 |
+
if "stop" in err_str and "400" in err_str and "stop" in kwargs:
|
| 190 |
+
kwargs.pop("stop", None)
|
| 191 |
+
try:
|
| 192 |
+
return original_create(*args, **kwargs)
|
| 193 |
+
except Exception as inner_e:
|
| 194 |
+
last_err = inner_e
|
| 195 |
+
err_str = str(inner_e).lower()
|
| 196 |
+
|
| 197 |
+
# Handle rate limit (429), payload/TPM limit (413), or server overload (503)
|
| 198 |
+
if any(code in err_str for code in ["429", "rate limit", "rate_limit", "413", "503", "404"]):
|
| 199 |
+
wait_time = 6 if len(verified_models) > 1 and attempt < len(verified_models) else 15 * (attempt + 1)
|
| 200 |
+
print(f"Step limit/error on {active_model} (attempt {attempt + 1}/5). Waiting {wait_time}s and switching model...")
|
| 201 |
+
time.sleep(wait_time)
|
| 202 |
+
else:
|
| 203 |
+
raise e
|
| 204 |
+
|
| 205 |
+
raise last_err
|
| 206 |
+
|
| 207 |
+
model.client.chat.completions.create = resilient_create
|
| 208 |
+
return model
|
| 209 |
|
| 210 |
def build_gaia_agent() -> CodeAgent:
|
| 211 |
+
"""Build CodeAgent using verified Groq models."""
|
| 212 |
groq_key = os.getenv("GROQ_API_KEY")
|
| 213 |
+
if not groq_key:
|
| 214 |
+
raise RuntimeError("GROQ_API_KEY secret is missing in Space settings.")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 215 |
|
| 216 |
+
model = create_resilient_groq_model(groq_key)
|
| 217 |
search_tool = DuckDuckGoSearchTool()
|
| 218 |
web_visit_tool = VisitWebpageTool()
|
| 219 |
|
|
|
|
| 246 |
"zipfile",
|
| 247 |
"base64"
|
| 248 |
],
|
| 249 |
+
max_steps=10
|
| 250 |
)
|
| 251 |
return agent
|
| 252 |
|
| 253 |
class GAIAgentWrapper:
|
| 254 |
+
"""Wrapper to run agent and enforce strict raw output normalization."""
|
| 255 |
def __init__(self):
|
| 256 |
print("Initializing CodeAgent...")
|
| 257 |
self.agent = build_gaia_agent()
|
|
|
|
| 269 |
"6. Never start with 'FINAL ANSWER:'. Output ONLY the pure value."
|
| 270 |
)
|
| 271 |
|
| 272 |
+
try:
|
| 273 |
+
raw_result = self.agent.run(strict_instruction)
|
| 274 |
+
cleaned = str(raw_result).strip()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 275 |
|
| 276 |
+
# Clean common prefixes
|
| 277 |
+
for prefix in ["final answer:", "answer:", "the answer is"]:
|
| 278 |
+
if cleaned.lower().startswith(prefix):
|
| 279 |
+
cleaned = cleaned[len(prefix):].strip(" :\n*`")
|
|
|
|
|
|
|
|
|
|
| 280 |
|
| 281 |
+
cleaned = cleaned.strip(" '\"`*")
|
| 282 |
|
| 283 |
+
# Normalize whole-number floats (e.g., '42.0' -> '42')
|
| 284 |
+
try:
|
| 285 |
+
as_float = float(cleaned)
|
| 286 |
+
if as_float == int(as_float):
|
| 287 |
+
cleaned = str(int(as_float))
|
| 288 |
+
except ValueError:
|
| 289 |
+
pass
|
|
|
|
|
|
|
| 290 |
|
| 291 |
+
return cleaned
|
| 292 |
+
except Exception as e:
|
| 293 |
+
print(f"Agent execution error: {e}")
|
| 294 |
+
return f"AGENT ERROR: {e}"
|
| 295 |
|
| 296 |
# --- File Handling ---
|
| 297 |
|
|
|
|
| 390 |
results_log = []
|
| 391 |
answers_payload = []
|
| 392 |
print(f"Running agent on {len(questions_data)} questions...")
|
| 393 |
+
for idx, item in enumerate(questions_data, 1):
|
| 394 |
task_id = item.get("task_id")
|
| 395 |
question_text = item.get("question")
|
| 396 |
suggested_filename = item.get("file_name")
|
| 397 |
if not task_id or question_text is None:
|
| 398 |
continue
|
| 399 |
|
| 400 |
+
print(f"--- Solving Question {idx}/{len(questions_data)} (Task {task_id}) ---")
|
| 401 |
file_path = download_task_file(api_url, task_id, suggested_name=suggested_filename)
|
| 402 |
if file_path:
|
| 403 |
file_ext = os.path.splitext(file_path)[-1]
|
|
|
|
| 417 |
print(f"Error running agent on task {task_id}: {e}")
|
| 418 |
results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"AGENT ERROR: {e}"})
|
| 419 |
|
| 420 |
+
time.sleep(3)
|
| 421 |
|
| 422 |
if not answers_payload:
|
| 423 |
return "Agent did not produce any answers to submit.", pd.DataFrame(results_log)
|