Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,236 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import streamlit as st
|
3 |
+
import openai
|
4 |
+
import pandas as pd
|
5 |
+
from typing import List, Tuple
|
6 |
+
from uuid import uuid4
|
7 |
+
import time
|
8 |
+
|
9 |
+
# π Set the OpenAI API key from an environment variable
|
10 |
+
openai.api_key = os.getenv("OPENAI_API_KEY")
|
11 |
+
|
12 |
+
# π Function to generate a unique session ID for caching
|
13 |
+
def get_session_id():
|
14 |
+
if 'session_id' not in st.session_state:
|
15 |
+
st.session_state.session_id = str(uuid4())
|
16 |
+
return st.session_state.session_id
|
17 |
+
|
18 |
+
# π§ STaR Algorithm Implementation
|
19 |
+
class SelfTaughtReasoner:
|
20 |
+
def __init__(self, model_engine="text-davinci-003"):
|
21 |
+
self.model_engine = model_engine
|
22 |
+
self.prompt_examples = [] # Initialize with an empty list
|
23 |
+
self.iterations = 0
|
24 |
+
self.generated_data = pd.DataFrame(columns=['Problem', 'Rationale', 'Answer', 'Is_Correct'])
|
25 |
+
self.rationalized_data = pd.DataFrame(columns=['Problem', 'Rationale', 'Answer', 'Is_Correct'])
|
26 |
+
self.fine_tuned_model = None # ποΈ Placeholder for fine-tuned model
|
27 |
+
|
28 |
+
def add_prompt_example(self, problem: str, rationale: str, answer: str):
|
29 |
+
"""
|
30 |
+
β Adds a prompt example to the few-shot examples.
|
31 |
+
"""
|
32 |
+
self.prompt_examples.append({
|
33 |
+
'Problem': problem,
|
34 |
+
'Rationale': rationale,
|
35 |
+
'Answer': answer
|
36 |
+
})
|
37 |
+
|
38 |
+
def construct_prompt(self, problem: str, include_answer: bool = False, answer: str = "") -> str:
|
39 |
+
"""
|
40 |
+
π Constructs the prompt for the OpenAI API call.
|
41 |
+
"""
|
42 |
+
prompt = ""
|
43 |
+
for example in self.prompt_examples:
|
44 |
+
prompt += f"Problem: {example['Problem']}\n"
|
45 |
+
prompt += f"Rationale: {example['Rationale']}\n"
|
46 |
+
prompt += f"Answer: {example['Answer']}\n\n"
|
47 |
+
|
48 |
+
prompt += f"Problem: {problem}\n"
|
49 |
+
if include_answer:
|
50 |
+
prompt += f"Answer (as hint): {answer}\n"
|
51 |
+
prompt += "Rationale:"
|
52 |
+
return prompt
|
53 |
+
|
54 |
+
def generate_rationale_and_answer(self, problem: str) -> Tuple[str, str]:
|
55 |
+
"""
|
56 |
+
π€ Generates a rationale and answer for a given problem.
|
57 |
+
"""
|
58 |
+
prompt = self.construct_prompt(problem)
|
59 |
+
try:
|
60 |
+
response = openai.Completion.create(
|
61 |
+
engine=self.model_engine,
|
62 |
+
prompt=prompt,
|
63 |
+
max_tokens=150,
|
64 |
+
temperature=0.7,
|
65 |
+
top_p=1,
|
66 |
+
frequency_penalty=0,
|
67 |
+
presence_penalty=0,
|
68 |
+
stop=["\n\n", "Problem:", "Answer:"]
|
69 |
+
)
|
70 |
+
rationale = response.choices[0].text.strip()
|
71 |
+
# π Now generate the answer using the rationale
|
72 |
+
prompt += f" {rationale}\nAnswer:"
|
73 |
+
answer_response = openai.Completion.create(
|
74 |
+
engine=self.model_engine,
|
75 |
+
prompt=prompt,
|
76 |
+
max_tokens=10,
|
77 |
+
temperature=0,
|
78 |
+
top_p=1,
|
79 |
+
frequency_penalty=0,
|
80 |
+
presence_penalty=0,
|
81 |
+
stop=["\n", "\n\n", "Problem:"]
|
82 |
+
)
|
83 |
+
answer = answer_response.choices[0].text.strip()
|
84 |
+
return rationale, answer
|
85 |
+
except Exception as e:
|
86 |
+
st.error(f"β Error generating rationale and answer: {e}")
|
87 |
+
return "", ""
|
88 |
+
|
89 |
+
def fine_tune_model(self):
|
90 |
+
"""
|
91 |
+
π οΈ Fine-tunes the model on the generated rationales.
|
92 |
+
"""
|
93 |
+
time.sleep(1) # β³ Simulate time taken for fine-tuning
|
94 |
+
self.fine_tuned_model = f"{self.model_engine}-fine-tuned-{get_session_id()}"
|
95 |
+
st.success(f"β
Model fine-tuned: {self.fine_tuned_model}")
|
96 |
+
|
97 |
+
def run_iteration(self, dataset: pd.DataFrame):
|
98 |
+
"""
|
99 |
+
π Runs one iteration of the STaR process.
|
100 |
+
"""
|
101 |
+
st.write(f"### Iteration {self.iterations + 1}")
|
102 |
+
progress_bar = st.progress(0)
|
103 |
+
total = len(dataset)
|
104 |
+
for idx, row in dataset.iterrows():
|
105 |
+
problem = row['Problem']
|
106 |
+
correct_answer = row['Answer']
|
107 |
+
# π€ Generate rationale and answer
|
108 |
+
rationale, answer = self.generate_rationale_and_answer(problem)
|
109 |
+
is_correct = (answer.lower() == correct_answer.lower())
|
110 |
+
# π Record the generated data
|
111 |
+
self.generated_data = self.generated_data.append({
|
112 |
+
'Problem': problem,
|
113 |
+
'Rationale': rationale,
|
114 |
+
'Answer': answer,
|
115 |
+
'Is_Correct': is_correct
|
116 |
+
}, ignore_index=True)
|
117 |
+
# β If incorrect, perform rationalization
|
118 |
+
if not is_correct:
|
119 |
+
rationale, answer = self.rationalize(problem, correct_answer)
|
120 |
+
is_correct = (answer.lower() == correct_answer.lower())
|
121 |
+
if is_correct:
|
122 |
+
self.rationalized_data = self.rationalized_data.append({
|
123 |
+
'Problem': problem,
|
124 |
+
'Rationale': rationale,
|
125 |
+
'Answer': answer,
|
126 |
+
'Is_Correct': is_correct
|
127 |
+
}, ignore_index=True)
|
128 |
+
progress_bar.progress((idx + 1) / total)
|
129 |
+
# π§ Fine-tune the model on correct rationales
|
130 |
+
st.write("π Fine-tuning the model on correct rationales...")
|
131 |
+
self.fine_tune_model()
|
132 |
+
self.iterations += 1
|
133 |
+
|
134 |
+
# Predefined problem and answer list for dataset
|
135 |
+
EXAMPLE_PROBLEM_ANSWERS = [
|
136 |
+
{"Problem": "What is deductive reasoning?", "Answer": "It is a logical process that draws specific conclusions from general principles."},
|
137 |
+
{"Problem": "What is inductive reasoning?", "Answer": "It is reasoning that forms general principles from specific examples."},
|
138 |
+
{"Problem": "Explain abductive reasoning.", "Answer": "It involves finding the best explanation for incomplete observations."},
|
139 |
+
{"Problem": "What is the capital of France?", "Answer": "Paris."},
|
140 |
+
{"Problem": "Who wrote Hamlet?", "Answer": "William Shakespeare."}
|
141 |
+
]
|
142 |
+
|
143 |
+
# Additional problem set for testing fine-tuned model
|
144 |
+
TEST_PROBLEM_SET = [
|
145 |
+
"What is the Pythagorean theorem?",
|
146 |
+
"Who developed the theory of relativity?",
|
147 |
+
"What is the main ingredient in bread?",
|
148 |
+
"Who is the author of 1984?",
|
149 |
+
"What is the boiling point of water?"
|
150 |
+
]
|
151 |
+
|
152 |
+
# Convert the example list into 'Problem | Answer' format
|
153 |
+
def format_examples_for_text_area(examples):
|
154 |
+
return '\n'.join([f"{example['Problem']} | {example['Answer']}" for example in examples])
|
155 |
+
|
156 |
+
# π₯οΈ Streamlit App
|
157 |
+
def main():
|
158 |
+
st.title("π€ Self-Taught Reasoner (STaR) Demonstration")
|
159 |
+
|
160 |
+
# π§© Initialize the Self-Taught Reasoner
|
161 |
+
if 'star' not in st.session_state:
|
162 |
+
st.session_state.star = SelfTaughtReasoner()
|
163 |
+
|
164 |
+
star = st.session_state.star
|
165 |
+
|
166 |
+
# Step 1: Few-Shot Prompt Examples
|
167 |
+
st.header("Step 1: Add Few-Shot Prompt Examples")
|
168 |
+
st.write("Choose an example from the dropdown or input your own.")
|
169 |
+
|
170 |
+
selected_example = st.selectbox(
|
171 |
+
"Select a predefined example",
|
172 |
+
[f"Example {i + 1}: {ex['Problem']}" for i, ex in enumerate(EXAMPLE_PROBLEM_ANSWERS)]
|
173 |
+
)
|
174 |
+
|
175 |
+
# Prefill with selected example
|
176 |
+
example_idx = int(selected_example.split(" ")[1].replace(":", "")) - 1
|
177 |
+
example_problem = EXAMPLE_PROBLEM_ANSWERS[example_idx]['Problem']
|
178 |
+
example_answer = EXAMPLE_PROBLEM_ANSWERS[example_idx]['Answer']
|
179 |
+
|
180 |
+
st.text_area("Problem", value=example_problem, height=50, key="example_problem")
|
181 |
+
st.text_input("Answer", value=example_answer, key="example_answer")
|
182 |
+
|
183 |
+
if st.button("Add Example"):
|
184 |
+
star.add_prompt_example(st.session_state.example_problem, "Rationale placeholder", st.session_state.example_answer)
|
185 |
+
st.success("Example added successfully!")
|
186 |
+
|
187 |
+
# Step 2: Input Dataset (Problem | Answer format)
|
188 |
+
st.header("Step 2: Input Dataset")
|
189 |
+
|
190 |
+
# Provide examples in the format 'Problem | Answer' as a default
|
191 |
+
prefilled_data = format_examples_for_text_area(EXAMPLE_PROBLEM_ANSWERS)
|
192 |
+
dataset_problems = st.text_area(
|
193 |
+
"Enter problems and answers in the format 'Problem | Answer', one per line.",
|
194 |
+
value=prefilled_data,
|
195 |
+
height=200
|
196 |
+
)
|
197 |
+
|
198 |
+
if st.button("Submit Dataset"):
|
199 |
+
dataset = []
|
200 |
+
lines = dataset_problems.strip().split('\n')
|
201 |
+
for line in lines:
|
202 |
+
if '|' in line:
|
203 |
+
problem, answer = line.split('|', 1)
|
204 |
+
dataset.append({'Problem': problem.strip(), 'Answer': answer.strip()})
|
205 |
+
st.session_state.dataset = pd.DataFrame(dataset)
|
206 |
+
st.success("Dataset loaded.")
|
207 |
+
|
208 |
+
if 'dataset' in st.session_state:
|
209 |
+
st.subheader("Current Dataset:")
|
210 |
+
st.dataframe(st.session_state.dataset.head())
|
211 |
+
|
212 |
+
# Step 3: Test the Fine-Tuned Model (renamed from Step 4)
|
213 |
+
st.header("Step 3: Test the Fine-Tuned Model")
|
214 |
+
|
215 |
+
# Add dropdown for selecting a test problem
|
216 |
+
test_problem = st.selectbox(
|
217 |
+
"Select a problem to test the fine-tuned model",
|
218 |
+
TEST_PROBLEM_SET
|
219 |
+
)
|
220 |
+
|
221 |
+
if st.button("Solve Problem"):
|
222 |
+
if not test_problem:
|
223 |
+
st.warning("Please enter or select a problem to solve.")
|
224 |
+
else:
|
225 |
+
rationale, answer = star.generate_rationale_and_answer(test_problem)
|
226 |
+
st.subheader("Rationale:")
|
227 |
+
st.write(rationale)
|
228 |
+
st.subheader("Answer:")
|
229 |
+
st.write(answer)
|
230 |
+
|
231 |
+
# Footer
|
232 |
+
st.write("---")
|
233 |
+
st.write("Developed as a demonstration of the STaR method.")
|
234 |
+
|
235 |
+
if __name__ == "__main__":
|
236 |
+
main()
|