Spaces:
Sleeping
Sleeping
Commit
·
b17581e
1
Parent(s):
cf3d2b0
update
Browse files
app.py
CHANGED
@@ -18,6 +18,55 @@ You are an AI tasked with identifying false negatives in answer verification. A
|
|
18 |
Return "True" if the model's answer is correct, otherwise return "False".
|
19 |
"""
|
20 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
# Main verification function
|
22 |
def verify_answer(question, ground_truth, model_answer, temperature, top_p, max_tokens):
|
23 |
# Format the prompt with user inputs
|
@@ -51,16 +100,28 @@ def verify_answer(question, ground_truth, model_answer, temperature, top_p, max_
|
|
51 |
except Exception as e:
|
52 |
yield f"Error: {str(e)}"
|
53 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
54 |
# Create the Gradio interface
|
55 |
-
with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue", font=[gr.themes.GoogleFont("Inter"), "ui-sans-serif", "system-ui", "sans-serif"]), title="
|
56 |
# Header with title and description
|
57 |
with gr.Row():
|
58 |
with gr.Column():
|
59 |
gr.Markdown(
|
60 |
"""
|
61 |
-
# Answer Verification Tool
|
62 |
|
63 |
-
This tool verifies if an answer is correct compared to a ground truth answer
|
64 |
"""
|
65 |
)
|
66 |
|
@@ -79,7 +140,7 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue", font=[gr.themes.GoogleFo
|
|
79 |
|
80 |
### What this tool does
|
81 |
|
82 |
-
This tool determines if a model's answer is semantically correct compared to a ground truth answer
|
83 |
|
84 |
The model analyzes both answers and returns:
|
85 |
- **True** if the model answer is correct
|
@@ -91,9 +152,9 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue", font=[gr.themes.GoogleFo
|
|
91 |
|
92 |
client = Client("zhangchenxu/TinyV")
|
93 |
result = client.predict(
|
94 |
-
question="
|
95 |
-
ground_truth="
|
96 |
-
model_answer="
|
97 |
temperature=0.3,
|
98 |
top_p=0.95,
|
99 |
max_tokens=1,
|
@@ -108,7 +169,7 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue", font=[gr.themes.GoogleFo
|
|
108 |
with gr.Accordion("Advanced Settings", open=False):
|
109 |
temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0.3, step=0.1, label="Temperature")
|
110 |
top_p = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)")
|
111 |
-
max_tokens = gr.Slider(minimum=
|
112 |
|
113 |
with gr.Column(scale=1):
|
114 |
gr.Markdown("## Input")
|
@@ -116,6 +177,19 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue", font=[gr.themes.GoogleFo
|
|
116 |
ground_truth = gr.Textbox(lines=5, label="Ground Truth Answer", placeholder="Enter the correct answer here...")
|
117 |
model_answer = gr.Textbox(lines=5, label="Model Answer", placeholder="Enter the answer to verify here...")
|
118 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
verify_btn = gr.Button("Verify Answer", variant="primary")
|
120 |
|
121 |
gr.Markdown("## Result")
|
@@ -128,68 +202,14 @@ with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue", font=[gr.themes.GoogleFo
|
|
128 |
outputs=result
|
129 |
)
|
130 |
|
131 |
-
#
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
"Paris is the capital of France.",
|
140 |
-
0.3,
|
141 |
-
0.95,
|
142 |
-
2,
|
143 |
-
],
|
144 |
-
[
|
145 |
-
"What is 2+2?",
|
146 |
-
"4",
|
147 |
-
"Four.",
|
148 |
-
0.3,
|
149 |
-
0.95,
|
150 |
-
2,
|
151 |
-
],
|
152 |
-
],
|
153 |
-
inputs=[question, ground_truth, model_answer, temperature, top_p, max_tokens],
|
154 |
-
outputs=result,
|
155 |
-
fn=verify_answer,
|
156 |
-
cache_examples=True,
|
157 |
-
)
|
158 |
-
|
159 |
-
with gr.Row():
|
160 |
-
gr.Examples(
|
161 |
-
examples=[
|
162 |
-
[
|
163 |
-
"When was the Declaration of Independence signed?",
|
164 |
-
"July 4, 1776",
|
165 |
-
"The Declaration of Independence was signed on July 4th, 1776.",
|
166 |
-
0.3,
|
167 |
-
0.95,
|
168 |
-
2,
|
169 |
-
],
|
170 |
-
[
|
171 |
-
"List the first three planets from the sun.",
|
172 |
-
"Mercury, Venus, Earth",
|
173 |
-
"The first three planets from the sun are Mercury, Venus, and Earth.",
|
174 |
-
0.3,
|
175 |
-
0.95,
|
176 |
-
2,
|
177 |
-
],
|
178 |
-
],
|
179 |
-
inputs=[question, ground_truth, model_answer, temperature, top_p, max_tokens],
|
180 |
-
outputs=result,
|
181 |
-
fn=verify_answer,
|
182 |
-
cache_examples=True,
|
183 |
-
)
|
184 |
-
|
185 |
-
# Add footer with extra information
|
186 |
-
with gr.Row(equal_height=True):
|
187 |
-
gr.Markdown(
|
188 |
-
"""
|
189 |
-
### About
|
190 |
-
This tool uses the zhangchenxu/TinyV-1.5B model to verify the correctness of the answers,
|
191 |
-
allowing for different phrasings of correct answers.
|
192 |
-
"""
|
193 |
)
|
194 |
|
195 |
# Define the public API
|
|
|
18 |
Return "True" if the model's answer is correct, otherwise return "False".
|
19 |
"""
|
20 |
|
21 |
+
# Define our example sets
|
22 |
+
EXAMPLES = [
|
23 |
+
{
|
24 |
+
"name": "Order-Insensitive",
|
25 |
+
"question": "Determine all real values of $x$ for which $(x+8)^{4}=(2 x+16)^{2}$.",
|
26 |
+
"ground_truth": "-6,-8,-10",
|
27 |
+
"model_answer": "-10, -8, -6",
|
28 |
+
"temp": 0.3,
|
29 |
+
"top_p": 0.95,
|
30 |
+
"tokens": 2
|
31 |
+
},
|
32 |
+
{
|
33 |
+
"name": "Latex Expression",
|
34 |
+
"question": "A bag contains 3 green balls, 4 red balls, and no other balls. Victor removes balls randomly from the bag, one at a time, and places them on a table. Each ball in the bag is equally likely to be chosen each time that he removes a ball. He stops removing balls when there are two balls of the same colour on the table. What is the probability that, when he stops, there is at least 1 red ball and at least 1 green ball on the table?",
|
35 |
+
"ground_truth": "$\\frac{4}{7}$",
|
36 |
+
"model_answer": "4/7",
|
37 |
+
"temp": 0.3,
|
38 |
+
"top_p": 0.95,
|
39 |
+
"tokens": 2
|
40 |
+
},
|
41 |
+
{
|
42 |
+
"name": "Variable Labeling",
|
43 |
+
"question": "If $T=x^{2}+\\frac{1}{x^{2}}$, determine the values of $b$ and $c$ so that $x^{6}+\\frac{1}{x^{6}}=T^{3}+b T+c$ for all non-zero real numbers $x$.",
|
44 |
+
"ground_truth": "-3,0",
|
45 |
+
"model_answer": "b=-3, c=0",
|
46 |
+
"temp": 0.3,
|
47 |
+
"top_p": 0.95,
|
48 |
+
"tokens": 2
|
49 |
+
},
|
50 |
+
{
|
51 |
+
"name": "Paraphrase",
|
52 |
+
"question": "Peter has 8 coins, of which he knows that 7 are genuine and weigh the same, while one is fake and differs in weight, though he does not know whether it is heavier or lighter. Peter has access to a balance scale, which shows which side is heavier but not by how much. For each weighing, Peter must pay Vasya one of his coins before the weighing. If Peter pays with a genuine coin, Vasya will provide an accurate result; if a fake coin is used, Vasya will provide a random result. Peter wants to determine 5 genuine coins and ensure that none of these genuine coins are given to Vasya. Can Peter guaranteedly achieve this?",
|
53 |
+
"ground_truth": "Petya can guarantee finding 5 genuine coins.",
|
54 |
+
"model_answer": "Yes, Peter can guarantee finding 5 genuine coins while ensuring that none of these genuine coins are paid to Vasya.",
|
55 |
+
"temp": 0.3,
|
56 |
+
"top_p": 0.95,
|
57 |
+
"tokens": 2
|
58 |
+
},
|
59 |
+
{
|
60 |
+
"name": "False Example",
|
61 |
+
"question": "What is the tallest mountain in the world?",
|
62 |
+
"ground_truth": "Mount Everest is the tallest mountain in the world.",
|
63 |
+
"model_answer": "K2 is the tallest mountain on Earth.",
|
64 |
+
"temp": 0.3,
|
65 |
+
"top_p": 0.95,
|
66 |
+
"tokens": 2
|
67 |
+
}
|
68 |
+
]
|
69 |
+
|
70 |
# Main verification function
|
71 |
def verify_answer(question, ground_truth, model_answer, temperature, top_p, max_tokens):
|
72 |
# Format the prompt with user inputs
|
|
|
100 |
except Exception as e:
|
101 |
yield f"Error: {str(e)}"
|
102 |
|
103 |
+
# Function to load an example when its button is clicked
|
104 |
+
def load_example(example_index):
|
105 |
+
example = EXAMPLES[example_index]
|
106 |
+
return (
|
107 |
+
example["question"],
|
108 |
+
example["ground_truth"],
|
109 |
+
example["model_answer"],
|
110 |
+
example["temp"],
|
111 |
+
example["top_p"],
|
112 |
+
example["tokens"]
|
113 |
+
)
|
114 |
+
|
115 |
# Create the Gradio interface
|
116 |
+
with gr.Blocks(theme=gr.themes.Soft(primary_hue="blue", font=[gr.themes.GoogleFont("Inter"), "ui-sans-serif", "system-ui", "sans-serif"]), title="TinyV") as demo:
|
117 |
# Header with title and description
|
118 |
with gr.Row():
|
119 |
with gr.Column():
|
120 |
gr.Markdown(
|
121 |
"""
|
122 |
+
# TinyV - Answer Verification Tool
|
123 |
|
124 |
+
This tool verifies if an answer is correct compared to a ground truth answer for RL.
|
125 |
"""
|
126 |
)
|
127 |
|
|
|
140 |
|
141 |
### What this tool does
|
142 |
|
143 |
+
This tool determines if a model's answer is semantically correct compared to a ground truth answer using a fine-tuned LLM.
|
144 |
|
145 |
The model analyzes both answers and returns:
|
146 |
- **True** if the model answer is correct
|
|
|
152 |
|
153 |
client = Client("zhangchenxu/TinyV")
|
154 |
result = client.predict(
|
155 |
+
question="Determine all real values of $x$ for which $(x+8)^{4}=(2 x+16)^{2}$.",
|
156 |
+
ground_truth="-6,-8,-10",
|
157 |
+
model_answer="-10, -8, -6",
|
158 |
temperature=0.3,
|
159 |
top_p=0.95,
|
160 |
max_tokens=1,
|
|
|
169 |
with gr.Accordion("Advanced Settings", open=False):
|
170 |
temperature = gr.Slider(minimum=0.0, maximum=1.0, value=0.3, step=0.1, label="Temperature")
|
171 |
top_p = gr.Slider(minimum=0.1, maximum=1.0, value=0.95, step=0.05, label="Top-p (nucleus sampling)")
|
172 |
+
max_tokens = gr.Slider(minimum=1, maximum=256, value=1, step=1, label="Max Tokens")
|
173 |
|
174 |
with gr.Column(scale=1):
|
175 |
gr.Markdown("## Input")
|
|
|
177 |
ground_truth = gr.Textbox(lines=5, label="Ground Truth Answer", placeholder="Enter the correct answer here...")
|
178 |
model_answer = gr.Textbox(lines=5, label="Model Answer", placeholder="Enter the answer to verify here...")
|
179 |
|
180 |
+
# Examples section as buttons
|
181 |
+
gr.Markdown("### Try an example:")
|
182 |
+
with gr.Row():
|
183 |
+
example_buttons = []
|
184 |
+
for i, example in enumerate(EXAMPLES):
|
185 |
+
btn = gr.Button(example["name"], size="sm")
|
186 |
+
example_buttons.append(btn)
|
187 |
+
# Connect each button to the load_example function
|
188 |
+
btn.click(
|
189 |
+
fn=lambda idx=i: load_example(idx),
|
190 |
+
outputs=[question, ground_truth, model_answer, temperature, top_p, max_tokens]
|
191 |
+
)
|
192 |
+
|
193 |
verify_btn = gr.Button("Verify Answer", variant="primary")
|
194 |
|
195 |
gr.Markdown("## Result")
|
|
|
202 |
outputs=result
|
203 |
)
|
204 |
|
205 |
+
# Run verification when an example is loaded (optional)
|
206 |
+
for btn in example_buttons:
|
207 |
+
btn.click(
|
208 |
+
fn=verify_answer,
|
209 |
+
inputs=[question, ground_truth, model_answer, temperature, top_p, max_tokens],
|
210 |
+
outputs=result,
|
211 |
+
_js="() => {setTimeout(() => document.querySelector('#verify-btn').click(), 100)}",
|
212 |
+
queue=False
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
213 |
)
|
214 |
|
215 |
# Define the public API
|