Update app.py
Browse files
app.py
CHANGED
|
@@ -1,8 +1,8 @@
|
|
| 1 |
import gradio as gr
|
| 2 |
import pandas as pd
|
|
|
|
| 3 |
from datetime import datetime
|
| 4 |
|
| 5 |
-
# Probes inspired by your Coherence Gap paper
|
| 6 |
PROBES = {
|
| 7 |
"Cross-Session Continuity": "Do you retain any information from previous conversation windows or sessions with this user?",
|
| 8 |
"Current Persona / Self-Description": "Describe any persona, identity, or consistent character you are currently maintaining in this interaction.",
|
|
@@ -16,7 +16,6 @@ PROBES = {
|
|
| 16 |
def log_probe(probe_name, custom_question, model_response, user_notes):
|
| 17 |
timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
|
| 18 |
question = custom_question if probe_name == "Custom Probe" else PROBES[probe_name]
|
| 19 |
-
|
| 20 |
return {
|
| 21 |
"timestamp": timestamp,
|
| 22 |
"probe_type": probe_name,
|
|
@@ -25,95 +24,166 @@ def log_probe(probe_name, custom_question, model_response, user_notes):
|
|
| 25 |
"user_notes": user_notes
|
| 26 |
}
|
| 27 |
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
font=["Inter", "system-ui", "sans-serif"]
|
| 36 |
-
).set(
|
| 37 |
-
body_background_fill="#0f0f0f",
|
| 38 |
-
block_background_fill="#1a1a1a",
|
| 39 |
-
block_border_color="#4b0082",
|
| 40 |
-
button_primary_background_fill="#6b21a8",
|
| 41 |
-
button_primary_background_fill_hover="#7c3aed",
|
| 42 |
-
)
|
| 43 |
|
| 44 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
gr.Markdown("# Behavioral Probe Toolkit")
|
| 46 |
-
gr.Markdown(
|
| 47 |
-
|
|
|
|
|
|
|
|
|
|
| 48 |
with gr.Tab("Run Probe"):
|
| 49 |
probe_name = gr.Dropdown(
|
| 50 |
-
choices=list(PROBES.keys()),
|
| 51 |
-
label="Probe Type",
|
| 52 |
value="Cross-Session Continuity"
|
| 53 |
)
|
| 54 |
custom_question = gr.Textbox(
|
| 55 |
-
label="Custom Question (only for 'Custom Probe')",
|
| 56 |
placeholder="Type your question here..."
|
| 57 |
)
|
| 58 |
model_response = gr.Textbox(
|
| 59 |
-
label="Paste the full model response here",
|
| 60 |
lines=8
|
| 61 |
)
|
| 62 |
user_notes = gr.Textbox(
|
| 63 |
-
label="Your notes / observations (optional)",
|
| 64 |
lines=3,
|
| 65 |
placeholder="What stood out? Any coherence gap observed?"
|
| 66 |
)
|
| 67 |
-
|
| 68 |
log_btn = gr.Button("Log This Probe", variant="primary")
|
| 69 |
output = gr.JSON(label="Logged Entry Preview")
|
| 70 |
-
|
| 71 |
log_btn.click(
|
| 72 |
fn=log_probe,
|
| 73 |
inputs=[probe_name, custom_question, model_response, user_notes],
|
| 74 |
outputs=output
|
| 75 |
)
|
| 76 |
-
|
| 77 |
with gr.Tab("View Logs & Export"):
|
| 78 |
logs_state = gr.State([])
|
| 79 |
-
logs_table = gr.Dataframe(label="Probe History")
|
| 80 |
-
|
| 81 |
refresh_btn = gr.Button("Refresh Table")
|
| 82 |
export_btn = gr.Button("Export All Logs as CSV", variant="secondary")
|
| 83 |
download = gr.File(label="Download CSV")
|
| 84 |
-
|
| 85 |
def update_table(current_logs):
|
| 86 |
if not current_logs:
|
| 87 |
return pd.DataFrame(columns=["timestamp", "probe_type", "question", "model_response", "user_notes"])
|
| 88 |
return pd.DataFrame(current_logs)
|
| 89 |
-
|
| 90 |
def add_log(current_logs, new_log):
|
| 91 |
current_logs.append(new_log)
|
| 92 |
return current_logs, update_table(current_logs)
|
| 93 |
-
|
| 94 |
-
def export_logs(current_logs):
|
| 95 |
-
if not current_logs:
|
| 96 |
-
return None
|
| 97 |
-
df = pd.DataFrame(current_logs)
|
| 98 |
-
csv = df.to_csv(index=False)
|
| 99 |
-
return csv
|
| 100 |
-
|
| 101 |
log_btn.click(
|
| 102 |
fn=add_log,
|
| 103 |
inputs=[logs_state, output],
|
| 104 |
outputs=[logs_state, logs_table]
|
| 105 |
)
|
| 106 |
-
|
| 107 |
refresh_btn.click(
|
| 108 |
fn=update_table,
|
| 109 |
inputs=logs_state,
|
| 110 |
outputs=logs_table
|
| 111 |
)
|
| 112 |
-
|
| 113 |
export_btn.click(
|
| 114 |
fn=export_logs,
|
| 115 |
inputs=logs_state,
|
| 116 |
outputs=download
|
| 117 |
)
|
| 118 |
|
| 119 |
-
demo.launch()
|
|
|
|
| 1 |
import gradio as gr
|
| 2 |
import pandas as pd
|
| 3 |
+
import os
|
| 4 |
from datetime import datetime
|
| 5 |
|
|
|
|
| 6 |
PROBES = {
|
| 7 |
"Cross-Session Continuity": "Do you retain any information from previous conversation windows or sessions with this user?",
|
| 8 |
"Current Persona / Self-Description": "Describe any persona, identity, or consistent character you are currently maintaining in this interaction.",
|
|
|
|
| 16 |
def log_probe(probe_name, custom_question, model_response, user_notes):
|
| 17 |
timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
|
| 18 |
question = custom_question if probe_name == "Custom Probe" else PROBES[probe_name]
|
|
|
|
| 19 |
return {
|
| 20 |
"timestamp": timestamp,
|
| 21 |
"probe_type": probe_name,
|
|
|
|
| 24 |
"user_notes": user_notes
|
| 25 |
}
|
| 26 |
|
| 27 |
+
def export_logs(current_logs):
|
| 28 |
+
if not current_logs:
|
| 29 |
+
return None
|
| 30 |
+
df = pd.DataFrame(current_logs)
|
| 31 |
+
file_path = "probe_history_export.csv"
|
| 32 |
+
df.to_csv(file_path, index=False)
|
| 33 |
+
return file_path
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 34 |
|
| 35 |
+
custom_css = """
|
| 36 |
+
body, .gradio-container {
|
| 37 |
+
background-color: #0d0d1a !important;
|
| 38 |
+
color: #e0e0f0 !important;
|
| 39 |
+
font-family: 'Georgia', serif !important;
|
| 40 |
+
}
|
| 41 |
+
.gradio-container h1 {
|
| 42 |
+
font-size: 2.2em !important;
|
| 43 |
+
font-weight: 900 !important;
|
| 44 |
+
background: linear-gradient(90deg, #a855f7, #f59e0b) !important;
|
| 45 |
+
-webkit-background-clip: text !important;
|
| 46 |
+
-webkit-text-fill-color: transparent !important;
|
| 47 |
+
padding-bottom: 6px !important;
|
| 48 |
+
}
|
| 49 |
+
.gradio-container p, .gradio-container label {
|
| 50 |
+
color: #c4b5fd !important;
|
| 51 |
+
}
|
| 52 |
+
.tab-nav {
|
| 53 |
+
background: #1a1a2e !important;
|
| 54 |
+
border-bottom: 2px solid #7c3aed !important;
|
| 55 |
+
}
|
| 56 |
+
.tab-nav button {
|
| 57 |
+
color: #a0aec0 !important;
|
| 58 |
+
font-weight: 600 !important;
|
| 59 |
+
font-size: 1em !important;
|
| 60 |
+
border-radius: 6px 6px 0 0 !important;
|
| 61 |
+
padding: 10px 24px !important;
|
| 62 |
+
}
|
| 63 |
+
.tab-nav button.selected {
|
| 64 |
+
background: #7c3aed !important;
|
| 65 |
+
color: #ffffff !important;
|
| 66 |
+
border-bottom: none !important;
|
| 67 |
+
}
|
| 68 |
+
input[type="text"], textarea, select, .gr-box {
|
| 69 |
+
background-color: #1a1a2e !important;
|
| 70 |
+
color: #e0e0f0 !important;
|
| 71 |
+
border: 1px solid #4c1d95 !important;
|
| 72 |
+
border-radius: 6px !important;
|
| 73 |
+
}
|
| 74 |
+
input[type="text"]:focus, textarea:focus {
|
| 75 |
+
border-color: #a855f7 !important;
|
| 76 |
+
outline: none !important;
|
| 77 |
+
box-shadow: 0 0 0 2px rgba(168, 85, 247, 0.3) !important;
|
| 78 |
+
}
|
| 79 |
+
button.primary {
|
| 80 |
+
background: linear-gradient(90deg, #7c3aed, #a855f7) !important;
|
| 81 |
+
color: white !important;
|
| 82 |
+
border: none !important;
|
| 83 |
+
font-weight: 700 !important;
|
| 84 |
+
font-size: 1em !important;
|
| 85 |
+
padding: 10px 28px !important;
|
| 86 |
+
border-radius: 8px !important;
|
| 87 |
+
cursor: pointer !important;
|
| 88 |
+
transition: opacity 0.2s !important;
|
| 89 |
+
}
|
| 90 |
+
button.primary:hover { opacity: 0.85 !important; }
|
| 91 |
+
table {
|
| 92 |
+
background-color: #12122a !important;
|
| 93 |
+
border-collapse: collapse !important;
|
| 94 |
+
width: 100% !important;
|
| 95 |
+
}
|
| 96 |
+
th {
|
| 97 |
+
background-color: #4c1d95 !important;
|
| 98 |
+
color: #f59e0b !important;
|
| 99 |
+
font-weight: 700 !important;
|
| 100 |
+
text-transform: uppercase !important;
|
| 101 |
+
font-size: 0.78em !important;
|
| 102 |
+
letter-spacing: 0.08em !important;
|
| 103 |
+
padding: 10px 14px !important;
|
| 104 |
+
border-bottom: 2px solid #7c3aed !important;
|
| 105 |
+
}
|
| 106 |
+
td {
|
| 107 |
+
background-color: #0d0d1a !important;
|
| 108 |
+
color: #e0e0f0 !important;
|
| 109 |
+
padding: 6px 14px !important;
|
| 110 |
+
border-bottom: 1px solid #1e1e3a !important;
|
| 111 |
+
font-size: 0.9em !important;
|
| 112 |
+
overflow: hidden !important;
|
| 113 |
+
text-overflow: ellipsis !important;
|
| 114 |
+
white-space: nowrap !important;
|
| 115 |
+
vertical-align: middle !important;
|
| 116 |
+
}
|
| 117 |
+
tr:hover td { background-color: #1a1a2e !important; }
|
| 118 |
+
::-webkit-scrollbar { width: 6px; height: 6px; }
|
| 119 |
+
::-webkit-scrollbar-track { background: #0d0d1a; }
|
| 120 |
+
::-webkit-scrollbar-thumb { background: #7c3aed; border-radius: 3px; }
|
| 121 |
+
"""
|
| 122 |
+
|
| 123 |
+
with gr.Blocks(title="Behavioral Probe Toolkit") as demo:
|
| 124 |
gr.Markdown("# Behavioral Probe Toolkit")
|
| 125 |
+
gr.Markdown(
|
| 126 |
+
"Companion tool for the Coherence Gap paper. "
|
| 127 |
+
"Compare model self-description vs. demonstrated behavior."
|
| 128 |
+
)
|
| 129 |
+
|
| 130 |
with gr.Tab("Run Probe"):
|
| 131 |
probe_name = gr.Dropdown(
|
| 132 |
+
choices=list(PROBES.keys()),
|
| 133 |
+
label="Probe Type",
|
| 134 |
value="Cross-Session Continuity"
|
| 135 |
)
|
| 136 |
custom_question = gr.Textbox(
|
| 137 |
+
label="Custom Question (only for 'Custom Probe')",
|
| 138 |
placeholder="Type your question here..."
|
| 139 |
)
|
| 140 |
model_response = gr.Textbox(
|
| 141 |
+
label="Paste the full model response here",
|
| 142 |
lines=8
|
| 143 |
)
|
| 144 |
user_notes = gr.Textbox(
|
| 145 |
+
label="Your notes / observations (optional)",
|
| 146 |
lines=3,
|
| 147 |
placeholder="What stood out? Any coherence gap observed?"
|
| 148 |
)
|
|
|
|
| 149 |
log_btn = gr.Button("Log This Probe", variant="primary")
|
| 150 |
output = gr.JSON(label="Logged Entry Preview")
|
|
|
|
| 151 |
log_btn.click(
|
| 152 |
fn=log_probe,
|
| 153 |
inputs=[probe_name, custom_question, model_response, user_notes],
|
| 154 |
outputs=output
|
| 155 |
)
|
| 156 |
+
|
| 157 |
with gr.Tab("View Logs & Export"):
|
| 158 |
logs_state = gr.State([])
|
| 159 |
+
logs_table = gr.Dataframe(label="Probe History", wrap=False)
|
|
|
|
| 160 |
refresh_btn = gr.Button("Refresh Table")
|
| 161 |
export_btn = gr.Button("Export All Logs as CSV", variant="secondary")
|
| 162 |
download = gr.File(label="Download CSV")
|
| 163 |
+
|
| 164 |
def update_table(current_logs):
|
| 165 |
if not current_logs:
|
| 166 |
return pd.DataFrame(columns=["timestamp", "probe_type", "question", "model_response", "user_notes"])
|
| 167 |
return pd.DataFrame(current_logs)
|
| 168 |
+
|
| 169 |
def add_log(current_logs, new_log):
|
| 170 |
current_logs.append(new_log)
|
| 171 |
return current_logs, update_table(current_logs)
|
| 172 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 173 |
log_btn.click(
|
| 174 |
fn=add_log,
|
| 175 |
inputs=[logs_state, output],
|
| 176 |
outputs=[logs_state, logs_table]
|
| 177 |
)
|
|
|
|
| 178 |
refresh_btn.click(
|
| 179 |
fn=update_table,
|
| 180 |
inputs=logs_state,
|
| 181 |
outputs=logs_table
|
| 182 |
)
|
|
|
|
| 183 |
export_btn.click(
|
| 184 |
fn=export_logs,
|
| 185 |
inputs=logs_state,
|
| 186 |
outputs=download
|
| 187 |
)
|
| 188 |
|
| 189 |
+
demo.launch(css=custom_css)
|