Spaces:
Runtime error
Runtime error
fixing
Browse files
app.py
CHANGED
|
@@ -1,307 +1,473 @@
|
|
| 1 |
import os
|
| 2 |
import gradio as gr
|
| 3 |
import requests
|
|
|
|
|
|
|
| 4 |
import json
|
| 5 |
import re
|
| 6 |
-
import
|
| 7 |
-
import
|
| 8 |
-
from
|
| 9 |
-
|
| 10 |
-
from typing import List, Dict, Optional, Tuple
|
| 11 |
-
from dotenv import load_dotenv
|
| 12 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 13 |
-
import torch
|
| 14 |
-
import time
|
| 15 |
-
import gc
|
| 16 |
|
| 17 |
-
#
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
MODEL_NAME = "microsoft/Phi-3-mini-4k-instruct"
|
| 21 |
-
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
| 22 |
|
| 23 |
# --- Constants ---
|
| 24 |
-
|
| 25 |
-
MAX_TOKENS = 256
|
| 26 |
-
TIMEOUT_PER_QUESTION = 45
|
| 27 |
-
MAX_RESULT_LENGTH = 500
|
| 28 |
-
MAX_ATTEMPTS = 2
|
| 29 |
-
|
| 30 |
-
# --- Model Initialization ---
|
| 31 |
-
print("Initializing model with fixed cache configuration...")
|
| 32 |
-
start_time = time.time()
|
| 33 |
-
|
| 34 |
-
model = AutoModelForCausalLM.from_pretrained(
|
| 35 |
-
MODEL_NAME,
|
| 36 |
-
trust_remote_code=True,
|
| 37 |
-
torch_dtype=torch.float32,
|
| 38 |
-
device_map="auto",
|
| 39 |
-
low_cpu_mem_usage=True
|
| 40 |
-
)
|
| 41 |
-
|
| 42 |
-
tokenizer = AutoTokenizer.from_pretrained(
|
| 43 |
-
MODEL_NAME,
|
| 44 |
-
use_fast=True,
|
| 45 |
-
trust_remote_code=True
|
| 46 |
-
)
|
| 47 |
-
|
| 48 |
-
if tokenizer.pad_token is None:
|
| 49 |
-
tokenizer.pad_token = tokenizer.eos_token
|
| 50 |
-
|
| 51 |
-
print(f"Model loaded in {time.time() - start_time:.2f} seconds")
|
| 52 |
|
| 53 |
-
|
| 54 |
-
def
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 69 |
|
| 70 |
-
|
| 71 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 72 |
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
def calculator(expression: str) -> str:
|
| 82 |
-
try:
|
| 83 |
-
expression = re.sub(r'[^\d+\-*/().^%,\s]', '', expression)
|
| 84 |
-
if not expression:
|
| 85 |
-
return "Invalid empty expression"
|
| 86 |
-
return str(numexpr.evaluate(expression))
|
| 87 |
-
except Exception as e:
|
| 88 |
-
return f"Calculation error: {str(e)}"
|
| 89 |
-
|
| 90 |
-
def read_webpage(url: str) -> str:
|
| 91 |
-
try:
|
| 92 |
-
if not re.match(r'^https?://', url):
|
| 93 |
-
return "Invalid URL format"
|
| 94 |
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 98 |
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
|
|
|
| 102 |
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
|
| 116 |
-
|
| 117 |
-
class GAIA_Agent:
|
| 118 |
-
def __init__(self):
|
| 119 |
-
self.tools = TOOLS
|
| 120 |
-
self.system_prompt = """You are an advanced problem solver. Follow these steps:
|
| 121 |
-
1. Analyze the question
|
| 122 |
-
2. Select the best tool
|
| 123 |
-
3. Execute with proper arguments
|
| 124 |
-
4. Interpret results
|
| 125 |
-
5. Provide final answer
|
| 126 |
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
|
|
|
|
|
|
| 131 |
|
| 132 |
-
|
| 133 |
-
{"tool": "tool_name", "args": {"arg": value}}```
|
| 134 |
|
| 135 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 136 |
|
| 137 |
-
|
| 138 |
-
start_time = time.time()
|
| 139 |
-
history = [f"Question: {question}"]
|
| 140 |
-
|
| 141 |
-
try:
|
| 142 |
-
for step in range(MAX_STEPS):
|
| 143 |
-
if time.time() - start_time > TIMEOUT_PER_QUESTION:
|
| 144 |
-
return "Timeout: Processing took too long"
|
| 145 |
-
|
| 146 |
-
prompt = self._build_prompt(history)
|
| 147 |
-
response = self._call_model(prompt)
|
| 148 |
-
|
| 149 |
-
if "Final Answer:" in response:
|
| 150 |
-
return response.split("Final Answer:")[-1].strip()[:500]
|
| 151 |
-
|
| 152 |
-
tool_call = self._parse_tool_call(response)
|
| 153 |
-
if tool_call:
|
| 154 |
-
tool_name, args = tool_call
|
| 155 |
-
observation = self._use_tool(tool_name, args)
|
| 156 |
-
history.append(f"Tool: {tool_name}")
|
| 157 |
-
history.append(f"Result: {observation[:300]}...")
|
| 158 |
-
else:
|
| 159 |
-
history.append(f"Thought: {response}")
|
| 160 |
-
|
| 161 |
-
gc.collect()
|
| 162 |
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
|
| 166 |
|
| 167 |
-
|
| 168 |
-
return f"<|system|>\n{self.system_prompt}<|end|>\n<|user|>\n" + "\n".join(history) + "<|end|>\n<|assistant|>"
|
| 169 |
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
prompt,
|
| 175 |
-
return_tensors="pt",
|
| 176 |
-
truncation=True,
|
| 177 |
-
max_length=3072,
|
| 178 |
-
padding=False
|
| 179 |
-
)
|
| 180 |
-
|
| 181 |
-
outputs = model.generate(
|
| 182 |
-
inputs.input_ids,
|
| 183 |
-
max_new_tokens=MAX_TOKENS,
|
| 184 |
-
temperature=0.3,
|
| 185 |
-
top_p=0.9,
|
| 186 |
-
do_sample=True,
|
| 187 |
-
pad_token_id=tokenizer.pad_token_id,
|
| 188 |
-
attention_mask=inputs.attention_mask
|
| 189 |
-
)
|
| 190 |
-
|
| 191 |
-
return tokenizer.decode(outputs[0], skip_special_tokens=True).split("<|assistant|>")[-1].strip()
|
| 192 |
-
except Exception as e:
|
| 193 |
-
if attempt < MAX_ATTEMPTS - 1:
|
| 194 |
-
time.sleep(0.5)
|
| 195 |
-
continue
|
| 196 |
-
return f"Model error: {str(e)}"
|
| 197 |
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
return None
|
| 207 |
-
if "tool" not in tool_call or "args" not in tool_call:
|
| 208 |
-
return None
|
| 209 |
-
if not isinstance(tool_call["args"], dict):
|
| 210 |
-
return None
|
| 211 |
-
|
| 212 |
-
return tool_call["tool"], tool_call["args"]
|
| 213 |
-
except:
|
| 214 |
-
return None
|
| 215 |
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
return f"Unknown tool: {tool_name}"
|
| 219 |
|
|
|
|
| 220 |
try:
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 227 |
|
| 228 |
-
return str(self.tools[tool_name](**args))[:MAX_RESULT_LENGTH]
|
| 229 |
except Exception as e:
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
def run_evaluation(profile: gr.OAuthProfile | None):
|
| 234 |
-
if not profile:
|
| 235 |
-
return "Please login first", None
|
| 236 |
-
|
| 237 |
-
agent = GAIA_Agent()
|
| 238 |
-
questions_url = f"{DEFAULT_API_URL}/questions"
|
| 239 |
-
submit_url = f"{DEFAULT_API_URL}/submit"
|
| 240 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 241 |
try:
|
| 242 |
-
response = requests.get(questions_url, timeout=
|
| 243 |
response.raise_for_status()
|
| 244 |
questions_data = response.json()
|
| 245 |
if not questions_data:
|
| 246 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 247 |
except Exception as e:
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
|
|
|
|
|
|
|
|
|
| 252 |
|
| 253 |
-
for i, item in enumerate(questions_data):
|
| 254 |
task_id = item.get("task_id")
|
| 255 |
-
|
| 256 |
|
| 257 |
-
if not task_id or
|
|
|
|
| 258 |
continue
|
| 259 |
|
| 260 |
-
print(f"Processing question {i
|
| 261 |
-
answer = agent(question)
|
| 262 |
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
"
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
|
| 273 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 274 |
}
|
| 275 |
|
|
|
|
|
|
|
|
|
|
| 276 |
try:
|
| 277 |
-
response = requests.post(submit_url, json=
|
| 278 |
response.raise_for_status()
|
| 279 |
-
|
| 280 |
-
|
| 281 |
-
|
| 282 |
-
|
| 283 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 284 |
except Exception as e:
|
| 285 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 286 |
|
| 287 |
-
# --- Gradio Interface ---
|
| 288 |
-
with gr.Blocks(title="Fixed GAIA Agent", theme=gr.themes.Soft()) as demo:
|
| 289 |
-
gr.Markdown("# ๐ GAIA Agent Evaluation")
|
| 290 |
-
|
| 291 |
with gr.Row():
|
| 292 |
gr.LoginButton()
|
| 293 |
-
|
| 294 |
-
|
| 295 |
-
|
| 296 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 297 |
|
| 298 |
-
|
| 299 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 300 |
outputs=[status_output, results_table]
|
| 301 |
)
|
| 302 |
|
| 303 |
if __name__ == "__main__":
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
|
| 307 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import os
|
| 2 |
import gradio as gr
|
| 3 |
import requests
|
| 4 |
+
import inspect
|
| 5 |
+
import pandas as pd
|
| 6 |
import json
|
| 7 |
import re
|
| 8 |
+
from typing import Dict, List, Any, Optional
|
| 9 |
+
import urllib.parse
|
| 10 |
+
from datetime import datetime
|
| 11 |
+
import math
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 12 |
|
| 13 |
+
# Transformers and torch imports
|
| 14 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
|
| 15 |
+
import torch
|
|
|
|
|
|
|
| 16 |
|
| 17 |
# --- Constants ---
|
| 18 |
+
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
+
class EnhancedGAIAAgent:
|
| 21 |
+
def __init__(self):
|
| 22 |
+
print("Initializing Enhanced GAIA Agent with Mistral-7B...")
|
| 23 |
+
|
| 24 |
+
# Initialize Mistral model
|
| 25 |
+
try:
|
| 26 |
+
print("Loading Mistral-7B-Instruct model...")
|
| 27 |
+
self.tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.3")
|
| 28 |
+
self.model = AutoModelForCausalLM.from_pretrained(
|
| 29 |
+
"mistralai/Mistral-7B-Instruct-v0.3",
|
| 30 |
+
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
|
| 31 |
+
device_map="auto" if torch.cuda.is_available() else None
|
| 32 |
+
)
|
| 33 |
|
| 34 |
+
# Create pipeline for easier use
|
| 35 |
+
self.pipe = pipeline(
|
| 36 |
+
"text-generation",
|
| 37 |
+
model=self.model,
|
| 38 |
+
tokenizer=self.tokenizer,
|
| 39 |
+
max_new_tokens=512,
|
| 40 |
+
temperature=0.7,
|
| 41 |
+
do_sample=True,
|
| 42 |
+
pad_token_id=self.tokenizer.eos_token_id
|
| 43 |
+
)
|
| 44 |
+
print("โ
Mistral model loaded successfully!")
|
| 45 |
+
|
| 46 |
+
except Exception as e:
|
| 47 |
+
print(f"โ Error loading Mistral model: {e}")
|
| 48 |
+
print("Falling back to basic responses...")
|
| 49 |
+
self.pipe = None
|
| 50 |
+
|
| 51 |
+
# Tool functions for GAIA tasks
|
| 52 |
+
self.tools = {
|
| 53 |
+
"calculate": self._calculate,
|
| 54 |
+
"search_web": self._search_web,
|
| 55 |
+
"parse_data": self._parse_data,
|
| 56 |
+
"analyze_text": self._analyze_text,
|
| 57 |
+
"solve_math": self._solve_math
|
| 58 |
+
}
|
| 59 |
|
| 60 |
+
def _calculate(self, expression: str) -> str:
|
| 61 |
+
"""Safe calculator for mathematical expressions"""
|
| 62 |
+
try:
|
| 63 |
+
# Clean and validate expression
|
| 64 |
+
expression = re.sub(r'[^0-9+\-*/().\s]', '', expression)
|
| 65 |
+
result = eval(expression)
|
| 66 |
+
return str(result)
|
| 67 |
+
except Exception as e:
|
| 68 |
+
return f"Calculation error: {e}"
|
| 69 |
+
|
| 70 |
+
def _search_web(self, query: str) -> str:
|
| 71 |
+
"""Simulate web search (placeholder - you'd integrate real search API)"""
|
| 72 |
+
# This is a placeholder - integrate with actual search API
|
| 73 |
+
return f"Search results for '{query}': [This would contain real search results]"
|
| 74 |
+
|
| 75 |
+
def _parse_data(self, data: str) -> str:
|
| 76 |
+
"""Parse and analyze structured data"""
|
| 77 |
+
try:
|
| 78 |
+
# Try to parse as JSON
|
| 79 |
+
if data.strip().startswith('{') or data.strip().startswith('['):
|
| 80 |
+
parsed = json.loads(data)
|
| 81 |
+
return f"Parsed data structure with {len(parsed) if isinstance(parsed, (list, dict)) else 1} elements"
|
| 82 |
+
else:
|
| 83 |
+
# Basic text analysis
|
| 84 |
+
lines = data.split('\n')
|
| 85 |
+
return f"Text data with {len(lines)} lines, {len(data.split())} words"
|
| 86 |
+
except Exception as e:
|
| 87 |
+
return f"Data parsing error: {e}"
|
| 88 |
+
|
| 89 |
+
def _analyze_text(self, text: str) -> str:
|
| 90 |
+
"""Analyze text content"""
|
| 91 |
+
words = text.split()
|
| 92 |
+
sentences = text.split('.')
|
| 93 |
+
return f"Text analysis: {len(words)} words, {len(sentences)} sentences"
|
| 94 |
+
|
| 95 |
+
def _solve_math(self, problem: str) -> str:
|
| 96 |
+
"""Enhanced math problem solver"""
|
| 97 |
+
try:
|
| 98 |
+
# Extract numbers and operations
|
| 99 |
+
numbers = re.findall(r'-?\d+\.?\d*', problem)
|
| 100 |
|
| 101 |
+
# Handle common math patterns
|
| 102 |
+
if "percent" in problem.lower() or "%" in problem:
|
| 103 |
+
if len(numbers) >= 2:
|
| 104 |
+
base = float(numbers[0])
|
| 105 |
+
percent = float(numbers[1])
|
| 106 |
+
result = base * (percent / 100)
|
| 107 |
+
return str(result)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 108 |
|
| 109 |
+
if "average" in problem.lower() or "mean" in problem.lower():
|
| 110 |
+
if numbers:
|
| 111 |
+
nums = [float(n) for n in numbers]
|
| 112 |
+
return str(sum(nums) / len(nums))
|
| 113 |
+
|
| 114 |
+
# Default calculation
|
| 115 |
+
return self._calculate(" ".join(numbers))
|
| 116 |
+
|
| 117 |
+
except Exception as e:
|
| 118 |
+
return f"Math solving error: {e}"
|
| 119 |
+
|
| 120 |
+
def _generate_response(self, prompt: str) -> str:
|
| 121 |
+
"""Generate response using Mistral model"""
|
| 122 |
+
if not self.pipe:
|
| 123 |
+
return "Model not available - using fallback response."
|
| 124 |
|
| 125 |
+
try:
|
| 126 |
+
messages = [
|
| 127 |
+
{"role": "user", "content": prompt}
|
| 128 |
+
]
|
| 129 |
|
| 130 |
+
response = self.pipe(messages, max_new_tokens=512, temperature=0.7)
|
| 131 |
+
|
| 132 |
+
# Extract the generated text
|
| 133 |
+
if response and len(response) > 0:
|
| 134 |
+
generated_text = response[0]['generated_text']
|
| 135 |
+
# Get only the assistant's response (after the user message)
|
| 136 |
+
if isinstance(generated_text, list):
|
| 137 |
+
# Find the assistant's response
|
| 138 |
+
for msg in generated_text:
|
| 139 |
+
if msg.get('role') == 'assistant':
|
| 140 |
+
return msg.get('content', '')
|
| 141 |
+
elif isinstance(generated_text, str):
|
| 142 |
+
return generated_text
|
| 143 |
+
else:
|
| 144 |
+
return str(generated_text)
|
| 145 |
+
|
| 146 |
+
return "No response generated."
|
| 147 |
+
|
| 148 |
+
except Exception as e:
|
| 149 |
+
print(f"Error generating response: {e}")
|
| 150 |
+
return f"Error in response generation: {e}"
|
| 151 |
+
|
| 152 |
+
def _detect_task_type(self, question: str) -> str:
|
| 153 |
+
"""Detect the type of task to apply appropriate strategy"""
|
| 154 |
+
question_lower = question.lower()
|
| 155 |
+
|
| 156 |
+
if any(word in question_lower for word in ["calculate", "compute", "math", "+", "-", "*", "/", "="]):
|
| 157 |
+
return "calculation"
|
| 158 |
+
elif any(word in question_lower for word in ["search", "find", "lookup", "google"]):
|
| 159 |
+
return "search"
|
| 160 |
+
elif any(word in question_lower for word in ["data", "csv", "json", "table", "parse"]):
|
| 161 |
+
return "data_analysis"
|
| 162 |
+
elif any(word in question_lower for word in ["percent", "%", "average", "mean", "sum"]):
|
| 163 |
+
return "math_word_problem"
|
| 164 |
+
else:
|
| 165 |
+
return "general_reasoning"
|
| 166 |
+
|
| 167 |
+
def __call__(self, question: str) -> str:
|
| 168 |
+
print(f"Agent processing question (first 100 chars): {question[:100]}...")
|
| 169 |
+
|
| 170 |
+
# Detect task type
|
| 171 |
+
task_type = self._detect_task_type(question)
|
| 172 |
+
print(f"Detected task type: {task_type}")
|
| 173 |
+
|
| 174 |
+
# Build enhanced prompt based on task type
|
| 175 |
+
if task_type == "calculation":
|
| 176 |
+
enhanced_prompt = f"""
|
| 177 |
+
You are a precise mathematical assistant. Solve this step-by-step:
|
| 178 |
|
| 179 |
+
Question: {question}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 180 |
|
| 181 |
+
Provide a clear, accurate answer. If calculation is needed, show your work.
|
| 182 |
+
Answer:"""
|
| 183 |
+
|
| 184 |
+
elif task_type == "math_word_problem":
|
| 185 |
+
enhanced_prompt = f"""
|
| 186 |
+
You are solving a math word problem. Break it down step by step:
|
| 187 |
|
| 188 |
+
Question: {question}
|
|
|
|
| 189 |
|
| 190 |
+
Steps:
|
| 191 |
+
1. Identify what is being asked
|
| 192 |
+
2. Extract the relevant numbers
|
| 193 |
+
3. Determine the operation needed
|
| 194 |
+
4. Calculate the result
|
| 195 |
+
5. Provide the final answer
|
| 196 |
|
| 197 |
+
Answer:"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 198 |
|
| 199 |
+
elif task_type == "data_analysis":
|
| 200 |
+
enhanced_prompt = f"""
|
| 201 |
+
You are analyzing data. Approach this systematically:
|
| 202 |
|
| 203 |
+
Question: {question}
|
|
|
|
| 204 |
|
| 205 |
+
Consider:
|
| 206 |
+
- What type of data is involved?
|
| 207 |
+
- What analysis is needed?
|
| 208 |
+
- What tools or methods should be used?
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 209 |
|
| 210 |
+
Provide a clear, structured answer.
|
| 211 |
+
Answer:"""
|
| 212 |
+
|
| 213 |
+
else:
|
| 214 |
+
enhanced_prompt = f"""
|
| 215 |
+
You are a helpful assistant that provides accurate, well-reasoned answers.
|
| 216 |
+
|
| 217 |
+
Question: {question}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 218 |
|
| 219 |
+
Think through this step-by-step and provide a clear, comprehensive answer.
|
| 220 |
+
Answer:"""
|
|
|
|
| 221 |
|
| 222 |
+
# Generate response using the model
|
| 223 |
try:
|
| 224 |
+
response = self._generate_response(enhanced_prompt)
|
| 225 |
+
|
| 226 |
+
# Post-process response for specific task types
|
| 227 |
+
if task_type in ["calculation", "math_word_problem"]:
|
| 228 |
+
# Try to extract and verify any calculations
|
| 229 |
+
numbers_in_response = re.findall(r'-?\d+\.?\d*', response)
|
| 230 |
+
if numbers_in_response:
|
| 231 |
+
# Attempt to verify calculation if simple enough
|
| 232 |
+
pass
|
| 233 |
+
|
| 234 |
+
print(f"Agent returning response (first 100 chars): {response[:100]}...")
|
| 235 |
+
return response.strip()
|
| 236 |
|
|
|
|
| 237 |
except Exception as e:
|
| 238 |
+
print(f"Error in agent processing: {e}")
|
| 239 |
+
fallback_response = self._handle_fallback(question, task_type)
|
| 240 |
+
return fallback_response
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 241 |
|
| 242 |
+
def _handle_fallback(self, question: str, task_type: str) -> str:
|
| 243 |
+
"""Provide fallback responses when the main model fails"""
|
| 244 |
+
if task_type == "calculation":
|
| 245 |
+
# Try to extract and calculate simple expressions
|
| 246 |
+
try:
|
| 247 |
+
numbers = re.findall(r'-?\d+\.?\d*', question)
|
| 248 |
+
if len(numbers) >= 2:
|
| 249 |
+
if "+" in question:
|
| 250 |
+
result = sum(float(n) for n in numbers)
|
| 251 |
+
return f"The sum is {result}"
|
| 252 |
+
elif "*" in question or "multiply" in question.lower():
|
| 253 |
+
result = 1
|
| 254 |
+
for n in numbers:
|
| 255 |
+
result *= float(n)
|
| 256 |
+
return f"The product is {result}"
|
| 257 |
+
except:
|
| 258 |
+
pass
|
| 259 |
+
|
| 260 |
+
return f"I understand you're asking about: {question}. This appears to be a {task_type} task. Let me provide my best analysis based on the available information."
|
| 261 |
+
|
| 262 |
+
|
| 263 |
+
def run_and_submit_all(profile: gr.OAuthProfile | None):
|
| 264 |
+
"""
|
| 265 |
+
Fetches all questions, runs the EnhancedGAIAAgent on them, submits all answers,
|
| 266 |
+
and displays the results.
|
| 267 |
+
"""
|
| 268 |
+
# --- Determine HF Space Runtime URL and Repo URL ---
|
| 269 |
+
space_id = os.getenv("SPACE_ID")
|
| 270 |
+
|
| 271 |
+
if profile:
|
| 272 |
+
username = f"{profile.username}"
|
| 273 |
+
print(f"User logged in: {username}")
|
| 274 |
+
else:
|
| 275 |
+
print("User not logged in.")
|
| 276 |
+
return "Please Login to Hugging Face with the button.", None
|
| 277 |
+
|
| 278 |
+
api_url = DEFAULT_API_URL
|
| 279 |
+
questions_url = f"{api_url}/questions"
|
| 280 |
+
submit_url = f"{api_url}/submit"
|
| 281 |
+
|
| 282 |
+
# 1. Instantiate Enhanced Agent
|
| 283 |
+
try:
|
| 284 |
+
print("Initializing Enhanced GAIA Agent...")
|
| 285 |
+
agent = EnhancedGAIAAgent()
|
| 286 |
+
print("โ
Agent initialized successfully!")
|
| 287 |
+
except Exception as e:
|
| 288 |
+
print(f"โ Error instantiating agent: {e}")
|
| 289 |
+
return f"Error initializing agent: {e}", None
|
| 290 |
+
|
| 291 |
+
agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
|
| 292 |
+
print(f"Agent code URL: {agent_code}")
|
| 293 |
+
|
| 294 |
+
# 2. Fetch Questions
|
| 295 |
+
print(f"Fetching questions from: {questions_url}")
|
| 296 |
try:
|
| 297 |
+
response = requests.get(questions_url, timeout=15)
|
| 298 |
response.raise_for_status()
|
| 299 |
questions_data = response.json()
|
| 300 |
if not questions_data:
|
| 301 |
+
print("Fetched questions list is empty.")
|
| 302 |
+
return "Fetched questions list is empty or invalid format.", None
|
| 303 |
+
print(f"โ
Fetched {len(questions_data)} questions.")
|
| 304 |
+
except requests.exceptions.RequestException as e:
|
| 305 |
+
print(f"โ Error fetching questions: {e}")
|
| 306 |
+
return f"Error fetching questions: {e}", None
|
| 307 |
+
except requests.exceptions.JSONDecodeError as e:
|
| 308 |
+
print(f"โ Error decoding JSON response from questions endpoint: {e}")
|
| 309 |
+
return f"Error decoding server response for questions: {e}", None
|
| 310 |
except Exception as e:
|
| 311 |
+
print(f"โ An unexpected error occurred fetching questions: {e}")
|
| 312 |
+
return f"An unexpected error occurred fetching questions: {e}", None
|
| 313 |
+
|
| 314 |
+
# 3. Run Enhanced Agent
|
| 315 |
+
results_log = []
|
| 316 |
+
answers_payload = []
|
| 317 |
+
print(f"๐ Running enhanced agent on {len(questions_data)} questions...")
|
| 318 |
|
| 319 |
+
for i, item in enumerate(questions_data, 1):
|
| 320 |
task_id = item.get("task_id")
|
| 321 |
+
question_text = item.get("question")
|
| 322 |
|
| 323 |
+
if not task_id or question_text is None:
|
| 324 |
+
print(f"โ ๏ธ Skipping item with missing task_id or question: {item}")
|
| 325 |
continue
|
| 326 |
|
| 327 |
+
print(f"๐ Processing question {i}/{len(questions_data)} (ID: {task_id})")
|
|
|
|
| 328 |
|
| 329 |
+
try:
|
| 330 |
+
submitted_answer = agent(question_text)
|
| 331 |
+
answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
|
| 332 |
+
results_log.append({
|
| 333 |
+
"Task ID": task_id,
|
| 334 |
+
"Question": question_text[:200] + "..." if len(question_text) > 200 else question_text,
|
| 335 |
+
"Submitted Answer": submitted_answer[:300] + "..." if len(submitted_answer) > 300 else submitted_answer
|
| 336 |
+
})
|
| 337 |
+
print(f"โ
Completed question {i}")
|
| 338 |
+
|
| 339 |
+
except Exception as e:
|
| 340 |
+
print(f"โ Error running agent on task {task_id}: {e}")
|
| 341 |
+
error_response = f"AGENT ERROR: {e}"
|
| 342 |
+
answers_payload.append({"task_id": task_id, "submitted_answer": error_response})
|
| 343 |
+
results_log.append({
|
| 344 |
+
"Task ID": task_id,
|
| 345 |
+
"Question": question_text[:200] + "..." if len(question_text) > 200 else question_text,
|
| 346 |
+
"Submitted Answer": error_response
|
| 347 |
+
})
|
| 348 |
+
|
| 349 |
+
if not answers_payload:
|
| 350 |
+
print("โ Agent did not produce any answers to submit.")
|
| 351 |
+
return "Agent did not produce any answers to submit.", pd.DataFrame(results_log)
|
| 352 |
+
|
| 353 |
+
# 4. Prepare Submission
|
| 354 |
+
submission_data = {
|
| 355 |
+
"username": username.strip(),
|
| 356 |
+
"agent_code": agent_code,
|
| 357 |
+
"answers": answers_payload
|
| 358 |
}
|
| 359 |
|
| 360 |
+
print(f"๐ค Submitting {len(answers_payload)} answers for user '{username}'...")
|
| 361 |
+
|
| 362 |
+
# 5. Submit
|
| 363 |
try:
|
| 364 |
+
response = requests.post(submit_url, json=submission_data, timeout=120) # Increased timeout
|
| 365 |
response.raise_for_status()
|
| 366 |
+
result_data = response.json()
|
| 367 |
+
|
| 368 |
+
final_status = (
|
| 369 |
+
f"๐ Submission Successful!\n"
|
| 370 |
+
f"User: {result_data.get('username')}\n"
|
| 371 |
+
f"Overall Score: {result_data.get('score', 'N/A')}% "
|
| 372 |
+
f"({result_data.get('correct_count', '?')}/{result_data.get('total_attempted', '?')} correct)\n"
|
| 373 |
+
f"Message: {result_data.get('message', 'No message received.')}"
|
| 374 |
+
)
|
| 375 |
+
|
| 376 |
+
print("โ
Submission successful!")
|
| 377 |
+
results_df = pd.DataFrame(results_log)
|
| 378 |
+
return final_status, results_df
|
| 379 |
+
|
| 380 |
+
except requests.exceptions.HTTPError as e:
|
| 381 |
+
error_detail = f"Server responded with status {e.response.status_code}."
|
| 382 |
+
try:
|
| 383 |
+
error_json = e.response.json()
|
| 384 |
+
error_detail += f" Detail: {error_json.get('detail', e.response.text)}"
|
| 385 |
+
except requests.exceptions.JSONDecodeError:
|
| 386 |
+
error_detail += f" Response: {e.response.text[:500]}"
|
| 387 |
+
status_message = f"โ Submission Failed: {error_detail}"
|
| 388 |
+
print(status_message)
|
| 389 |
+
results_df = pd.DataFrame(results_log)
|
| 390 |
+
return status_message, results_df
|
| 391 |
+
|
| 392 |
except Exception as e:
|
| 393 |
+
status_message = f"โ An unexpected error occurred during submission: {e}"
|
| 394 |
+
print(status_message)
|
| 395 |
+
results_df = pd.DataFrame(results_log)
|
| 396 |
+
return status_message, results_df
|
| 397 |
+
|
| 398 |
+
|
| 399 |
+
# --- Build Gradio Interface using Blocks ---
|
| 400 |
+
with gr.Blocks(title="Enhanced GAIA Agent") as demo:
|
| 401 |
+
gr.Markdown("# ๐ Enhanced GAIA Agent with Mistral-7B")
|
| 402 |
+
gr.Markdown(
|
| 403 |
+
"""
|
| 404 |
+
**Enhanced Features:**
|
| 405 |
+
- ๐ง **Mistral-7B-Instruct** for advanced reasoning
|
| 406 |
+
- ๐ง **Tool Integration** for calculations and data processing
|
| 407 |
+
- ๐ **Task Type Detection** for optimized responses
|
| 408 |
+
- ๐ฏ **GAIA-Optimized** prompting strategies
|
| 409 |
+
|
| 410 |
+
**Instructions:**
|
| 411 |
+
1. Clone this space and ensure you have access to Mistral-7B-Instruct
|
| 412 |
+
2. Log in to your Hugging Face account using the button below
|
| 413 |
+
3. Click 'Run Enhanced Evaluation' to process all questions with the enhanced agent
|
| 414 |
+
|
| 415 |
+
**Note:** The enhanced agent uses Mistral-7B which requires significant computational resources.
|
| 416 |
+
Processing may take several minutes depending on the number of questions.
|
| 417 |
+
"""
|
| 418 |
+
)
|
| 419 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 420 |
with gr.Row():
|
| 421 |
gr.LoginButton()
|
| 422 |
+
|
| 423 |
+
with gr.Row():
|
| 424 |
+
run_button = gr.Button("๐ Run Enhanced Evaluation & Submit All Answers", variant="primary")
|
| 425 |
+
|
| 426 |
+
status_output = gr.Textbox(
|
| 427 |
+
label="๐ Run Status / Submission Result",
|
| 428 |
+
lines=8,
|
| 429 |
+
interactive=False
|
| 430 |
+
)
|
| 431 |
|
| 432 |
+
results_table = gr.DataFrame(
|
| 433 |
+
label="๐ Questions and Agent Answers",
|
| 434 |
+
wrap=True,
|
| 435 |
+
height=400
|
| 436 |
+
)
|
| 437 |
+
|
| 438 |
+
run_button.click(
|
| 439 |
+
fn=run_and_submit_all,
|
| 440 |
outputs=[status_output, results_table]
|
| 441 |
)
|
| 442 |
|
| 443 |
if __name__ == "__main__":
|
| 444 |
+
print("\n" + "="*50)
|
| 445 |
+
print("๐ ENHANCED GAIA AGENT STARTING")
|
| 446 |
+
print("="*50)
|
| 447 |
+
|
| 448 |
+
# Environment check
|
| 449 |
+
space_host = os.getenv("SPACE_HOST")
|
| 450 |
+
space_id = os.getenv("SPACE_ID")
|
| 451 |
+
|
| 452 |
+
if space_host:
|
| 453 |
+
print(f"โ
SPACE_HOST: {space_host}")
|
| 454 |
+
print(f"๐ Runtime URL: https://{space_host}.hf.space")
|
| 455 |
+
else:
|
| 456 |
+
print("โน๏ธ Running locally - SPACE_HOST not found")
|
| 457 |
+
|
| 458 |
+
if space_id:
|
| 459 |
+
print(f"โ
SPACE_ID: {space_id}")
|
| 460 |
+
print(f"๐ Repo URL: https://huggingface.co/spaces/{space_id}")
|
| 461 |
+
else:
|
| 462 |
+
print("โน๏ธ SPACE_ID not found")
|
| 463 |
+
|
| 464 |
+
# GPU/CPU check
|
| 465 |
+
if torch.cuda.is_available():
|
| 466 |
+
print(f"๐ฎ GPU Available: {torch.cuda.get_device_name()}")
|
| 467 |
+
print(f"๐พ GPU Memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB")
|
| 468 |
+
else:
|
| 469 |
+
print("๐ป Running on CPU (GPU not available)")
|
| 470 |
+
|
| 471 |
+
print("="*50)
|
| 472 |
+
print("๐ Launching Enhanced GAIA Agent Interface...")
|
| 473 |
+
demo.launch(debug=True, share=False)
|