Update app.py
Browse files
app.py
CHANGED
|
@@ -9,7 +9,7 @@ import re
|
|
| 9 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
| 10 |
|
| 11 |
# --- Enhanced Agent ---
|
| 12 |
-
class EnhancedAgent:
|
| 13 |
def __init__(self):
|
| 14 |
print("Initializing Enhanced Agent...")
|
| 15 |
|
|
@@ -25,7 +25,7 @@ class EnhancedAgent:
|
|
| 25 |
from duckduckgo_search import DDGS
|
| 26 |
self.search = DDGS()
|
| 27 |
print("β
Search ready")
|
| 28 |
-
except:
|
| 29 |
self.search = None
|
| 30 |
print("β οΈ Search unavailable")
|
| 31 |
|
|
@@ -34,7 +34,7 @@ class EnhancedAgent:
|
|
| 34 |
if not self.search:
|
| 35 |
return ""
|
| 36 |
|
| 37 |
-
try:
|
| 38 |
results = list(self.search.text(query, max_results=max_results))
|
| 39 |
if not results:
|
| 40 |
return ""
|
|
@@ -58,24 +58,24 @@ class EnhancedAgent:
|
|
| 58 |
patterns_to_remove = [
|
| 59 |
r"^according to.*? [,:]",
|
| 60 |
r"^based on.*?[,:]",
|
| 61 |
-
r"^the answer is: ? ",
|
| 62 |
-
r"^answer: ?",
|
| 63 |
-
r"^final answer:?",
|
| 64 |
-
r"^in summary:?",
|
| 65 |
-
r"^therefore:?",
|
| 66 |
-
r"^thus:?",
|
| 67 |
-
r"^so:?",
|
| 68 |
]
|
| 69 |
|
| 70 |
cleaned = text. strip()
|
| 71 |
-
for pattern in patterns_to_remove:
|
| 72 |
cleaned = re.sub(pattern, "", cleaned, flags=re.IGNORECASE).strip()
|
| 73 |
|
| 74 |
# If answer has multiple sentences, often the last one is the direct answer
|
| 75 |
sentences = cleaned.split('.')
|
| 76 |
if len(sentences) > 2:
|
| 77 |
# Check if last sentence looks like a direct answer
|
| 78 |
-
last = sentences[-1].
|
| 79 |
if last and len(last) < 100:
|
| 80 |
return last
|
| 81 |
|
|
@@ -86,7 +86,7 @@ class EnhancedAgent:
|
|
| 86 |
print(f"Q: {question[: 150]}")
|
| 87 |
|
| 88 |
# Determine if we need search
|
| 89 |
-
needs_search = any(keyword in question.lower() for keyword in [
|
| 90 |
'current', 'latest', 'recent', 'today', 'now', '2024', '2025', '2026',
|
| 91 |
'who is', 'what is', 'where is', 'when did', 'how many'
|
| 92 |
])
|
|
@@ -99,7 +99,7 @@ class EnhancedAgent:
|
|
| 99 |
print(f"β
Search: {len(search_context)} chars")
|
| 100 |
|
| 101 |
# Enhanced system prompt with better instructions
|
| 102 |
-
system_prompt = """You are an expert AI that provides accurate, direct answers.
|
| 103 |
|
| 104 |
CRITICAL RULES:
|
| 105 |
1. Give ONLY the final answer - no explanations unless asked
|
|
@@ -119,7 +119,7 @@ Q: "Who is the CEO of Tesla?" β A: "Elon Musk"
|
|
| 119 |
Q: "What is 15 + 27?" β A: "42"
|
| 120 |
Q: "How many planets in the solar system?" β A: "8"
|
| 121 |
|
| 122 |
-
DO NOT start with "The answer is" or "According to" - just give the answer directly!
|
| 123 |
|
| 124 |
# Build prompt
|
| 125 |
messages = [{"role": "system", "content": system_prompt}]
|
|
@@ -146,7 +146,7 @@ DO NOT start with "The answer is" or "According to" - just give the answer direc
|
|
| 146 |
temperature=0.05 # Very low for maximum accuracy
|
| 147 |
)
|
| 148 |
|
| 149 |
-
raw_answer = response.choices[0].message.content.strip()
|
| 150 |
|
| 151 |
# Clean up the answer
|
| 152 |
answer = self.extract_answer(raw_answer)
|
|
@@ -160,7 +160,7 @@ DO NOT start with "The answer is" or "According to" - just give the answer direc
|
|
| 160 |
answer = line
|
| 161 |
break
|
| 162 |
|
| 163 |
-
print(f"β
A: {answer[:
|
| 164 |
return answer
|
| 165 |
|
| 166 |
except Exception as e:
|
|
@@ -169,14 +169,14 @@ DO NOT start with "The answer is" or "According to" - just give the answer direc
|
|
| 169 |
# Last resort: try a simpler call
|
| 170 |
try:
|
| 171 |
simple_prompt = f"Answer this question concisely:\n\n{question}\n\nAnswer:"
|
| 172 |
-
response = self.
|
| 173 |
simple_prompt,
|
| 174 |
model="mistralai/Mixtral-8x7B-Instruct-v0.1",
|
| 175 |
max_new_tokens=200,
|
| 176 |
temperature=0.1
|
| 177 |
)
|
| 178 |
return response.strip()
|
| 179 |
-
except:
|
| 180 |
return "Unable to generate answer."
|
| 181 |
|
| 182 |
return "Unable to generate answer."
|
|
@@ -229,7 +229,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
|
|
| 229 |
task_id = item.get("task_id")
|
| 230 |
question_text = item. get("question")
|
| 231 |
|
| 232 |
-
if not task_id or not question_text:
|
| 233 |
continue
|
| 234 |
|
| 235 |
print(f"[{idx}/{total}] {task_id[: 8]}...")
|
|
@@ -247,7 +247,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
|
|
| 247 |
"task_id": task_id,
|
| 248 |
"submitted_answer": "Error processing question"
|
| 249 |
})
|
| 250 |
-
results_log.append((idx, question_text[: 60], f"Error: {str(e)[:
|
| 251 |
|
| 252 |
print(f"\n{'='*70}")
|
| 253 |
print(f"β
Processed all {len(answers_payload)} questions")
|
|
@@ -284,7 +284,7 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
|
|
| 284 |
|
| 285 |
### π Score: **{score}%**
|
| 286 |
|
| 287 |
-
{'### π Amazing!
|
| 288 |
|
| 289 |
**Details:**
|
| 290 |
- π€ User: `{username}`
|
|
@@ -330,12 +330,11 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
|
|
| 330 |
submit_button = gr.Button(
|
| 331 |
"π Run Evaluation & Submit",
|
| 332 |
variant="primary",
|
| 333 |
-
size="lg"
|
| 334 |
-
scale=2
|
| 335 |
)
|
| 336 |
|
| 337 |
output_text = gr.Markdown()
|
| 338 |
-
output_table = gr.Dataframe(label="π Results Preview", wrap=True
|
| 339 |
|
| 340 |
submit_button.click(
|
| 341 |
run_and_submit_all,
|
|
|
|
| 9 |
DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
|
| 10 |
|
| 11 |
# --- Enhanced Agent ---
|
| 12 |
+
class EnhancedAgent:
|
| 13 |
def __init__(self):
|
| 14 |
print("Initializing Enhanced Agent...")
|
| 15 |
|
|
|
|
| 25 |
from duckduckgo_search import DDGS
|
| 26 |
self.search = DDGS()
|
| 27 |
print("β
Search ready")
|
| 28 |
+
except:
|
| 29 |
self.search = None
|
| 30 |
print("β οΈ Search unavailable")
|
| 31 |
|
|
|
|
| 34 |
if not self.search:
|
| 35 |
return ""
|
| 36 |
|
| 37 |
+
try:
|
| 38 |
results = list(self.search.text(query, max_results=max_results))
|
| 39 |
if not results:
|
| 40 |
return ""
|
|
|
|
| 58 |
patterns_to_remove = [
|
| 59 |
r"^according to.*? [,:]",
|
| 60 |
r"^based on.*?[,:]",
|
| 61 |
+
r"^the answer is: ? ? ",
|
| 62 |
+
r"^answer: ? ?",
|
| 63 |
+
r"^final answer:? ?",
|
| 64 |
+
r"^in summary:? ?",
|
| 65 |
+
r"^therefore:? ?",
|
| 66 |
+
r"^thus:? ?",
|
| 67 |
+
r"^so:? ?",
|
| 68 |
]
|
| 69 |
|
| 70 |
cleaned = text. strip()
|
| 71 |
+
for pattern in patterns_to_remove:
|
| 72 |
cleaned = re.sub(pattern, "", cleaned, flags=re.IGNORECASE).strip()
|
| 73 |
|
| 74 |
# If answer has multiple sentences, often the last one is the direct answer
|
| 75 |
sentences = cleaned.split('.')
|
| 76 |
if len(sentences) > 2:
|
| 77 |
# Check if last sentence looks like a direct answer
|
| 78 |
+
last = sentences[-1].strip()
|
| 79 |
if last and len(last) < 100:
|
| 80 |
return last
|
| 81 |
|
|
|
|
| 86 |
print(f"Q: {question[: 150]}")
|
| 87 |
|
| 88 |
# Determine if we need search
|
| 89 |
+
needs_search = any(keyword in question. lower() for keyword in [
|
| 90 |
'current', 'latest', 'recent', 'today', 'now', '2024', '2025', '2026',
|
| 91 |
'who is', 'what is', 'where is', 'when did', 'how many'
|
| 92 |
])
|
|
|
|
| 99 |
print(f"β
Search: {len(search_context)} chars")
|
| 100 |
|
| 101 |
# Enhanced system prompt with better instructions
|
| 102 |
+
system_prompt = """You are an expert AI that provides accurate, direct answers.
|
| 103 |
|
| 104 |
CRITICAL RULES:
|
| 105 |
1. Give ONLY the final answer - no explanations unless asked
|
|
|
|
| 119 |
Q: "What is 15 + 27?" β A: "42"
|
| 120 |
Q: "How many planets in the solar system?" β A: "8"
|
| 121 |
|
| 122 |
+
DO NOT start with "The answer is" or "According to" - just give the answer directly!"""
|
| 123 |
|
| 124 |
# Build prompt
|
| 125 |
messages = [{"role": "system", "content": system_prompt}]
|
|
|
|
| 146 |
temperature=0.05 # Very low for maximum accuracy
|
| 147 |
)
|
| 148 |
|
| 149 |
+
raw_answer = response.choices[0].message.content. strip()
|
| 150 |
|
| 151 |
# Clean up the answer
|
| 152 |
answer = self.extract_answer(raw_answer)
|
|
|
|
| 160 |
answer = line
|
| 161 |
break
|
| 162 |
|
| 163 |
+
print(f"β
A: {answer[:200]}")
|
| 164 |
return answer
|
| 165 |
|
| 166 |
except Exception as e:
|
|
|
|
| 169 |
# Last resort: try a simpler call
|
| 170 |
try:
|
| 171 |
simple_prompt = f"Answer this question concisely:\n\n{question}\n\nAnswer:"
|
| 172 |
+
response = self.client.text_generation(
|
| 173 |
simple_prompt,
|
| 174 |
model="mistralai/Mixtral-8x7B-Instruct-v0.1",
|
| 175 |
max_new_tokens=200,
|
| 176 |
temperature=0.1
|
| 177 |
)
|
| 178 |
return response.strip()
|
| 179 |
+
except:
|
| 180 |
return "Unable to generate answer."
|
| 181 |
|
| 182 |
return "Unable to generate answer."
|
|
|
|
| 229 |
task_id = item.get("task_id")
|
| 230 |
question_text = item. get("question")
|
| 231 |
|
| 232 |
+
if not task_id or not question_text:
|
| 233 |
continue
|
| 234 |
|
| 235 |
print(f"[{idx}/{total}] {task_id[: 8]}...")
|
|
|
|
| 247 |
"task_id": task_id,
|
| 248 |
"submitted_answer": "Error processing question"
|
| 249 |
})
|
| 250 |
+
results_log.append((idx, question_text[: 60], f"Error: {str(e)[:50]}"))
|
| 251 |
|
| 252 |
print(f"\n{'='*70}")
|
| 253 |
print(f"β
Processed all {len(answers_payload)} questions")
|
|
|
|
| 284 |
|
| 285 |
### π Score: **{score}%**
|
| 286 |
|
| 287 |
+
{'### π Amazing! You completed Unit 4 of the Hugging Face Agents Course!' if passed else f'### π You got {score}% - need 30% to pass. The agent is now much better, try again!'}
|
| 288 |
|
| 289 |
**Details:**
|
| 290 |
- π€ User: `{username}`
|
|
|
|
| 330 |
submit_button = gr.Button(
|
| 331 |
"π Run Evaluation & Submit",
|
| 332 |
variant="primary",
|
| 333 |
+
size="lg"
|
|
|
|
| 334 |
)
|
| 335 |
|
| 336 |
output_text = gr.Markdown()
|
| 337 |
+
output_table = gr.Dataframe(label="π Results Preview", wrap=True)
|
| 338 |
|
| 339 |
submit_button.click(
|
| 340 |
run_and_submit_all,
|