Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -23,52 +23,65 @@ def parse_input(json_input):
|
|
| 23 |
logger.debug("Successfully parsed as JSON")
|
| 24 |
return data
|
| 25 |
except json.JSONDecodeError as e:
|
| 26 |
-
logger.error("JSON parsing failed: %s", str(e))
|
| 27 |
-
raise ValueError(f"Malformed JSON: {str(e)}. Use double quotes for property names (e.g., \"content\").")
|
| 28 |
|
| 29 |
# Function to ensure a value is a float
|
| 30 |
def ensure_float(value):
|
| 31 |
if value is None:
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
return float(value)
|
| 35 |
if isinstance(value, str):
|
| 36 |
try:
|
| 37 |
return float(value)
|
| 38 |
except ValueError:
|
| 39 |
-
logger.error("
|
| 40 |
-
return 0.0
|
| 41 |
-
|
|
|
|
|
|
|
| 42 |
|
| 43 |
-
# Function to get token value
|
| 44 |
def get_token(entry):
|
| 45 |
-
|
|
|
|
|
|
|
|
|
|
| 46 |
|
| 47 |
# Function to create an empty Plotly figure
|
| 48 |
def create_empty_figure(title):
|
| 49 |
return go.Figure().update_layout(title=title, xaxis_title="", yaxis_title="", showlegend=False)
|
| 50 |
|
| 51 |
-
#
|
| 52 |
async def precompute_chunk(json_input, chunk_size, current_chunk):
|
| 53 |
try:
|
| 54 |
data = parse_input(json_input)
|
| 55 |
content = data.get("content", []) if isinstance(data, dict) else data
|
| 56 |
if not isinstance(content, list):
|
| 57 |
-
raise ValueError("Content must be a list")
|
| 58 |
|
| 59 |
tokens = []
|
| 60 |
logprobs = []
|
| 61 |
top_alternatives = []
|
| 62 |
for entry in content:
|
| 63 |
if not isinstance(entry, dict):
|
|
|
|
| 64 |
continue
|
| 65 |
logprob = ensure_float(entry.get("logprob", None))
|
| 66 |
-
if logprob >= -100000:
|
| 67 |
tokens.append(get_token(entry))
|
| 68 |
logprobs.append(logprob)
|
| 69 |
-
top_probs = entry.get("top_logprobs", {})
|
| 70 |
-
|
| 71 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 72 |
|
| 73 |
if not tokens or not logprobs:
|
| 74 |
return None, None, None
|
|
@@ -79,7 +92,11 @@ async def precompute_chunk(json_input, chunk_size, current_chunk):
|
|
| 79 |
if start_idx >= len(tokens):
|
| 80 |
return None, None, None
|
| 81 |
|
| 82 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
except Exception as e:
|
| 84 |
logger.error("Precomputation failed for chunk %d: %s", current_chunk + 1, str(e))
|
| 85 |
return None, None, None
|
|
@@ -97,146 +114,402 @@ def precompute_next_chunk_sync(json_input, current_chunk):
|
|
| 97 |
loop.close()
|
| 98 |
return result
|
| 99 |
|
| 100 |
-
#
|
| 101 |
def visualize_logprobs(json_input, chunk=0, chunk_size=100):
|
| 102 |
try:
|
| 103 |
data = parse_input(json_input)
|
| 104 |
content = data.get("content", []) if isinstance(data, dict) else data
|
| 105 |
if not isinstance(content, list):
|
| 106 |
-
raise ValueError("Content must be a list")
|
| 107 |
|
| 108 |
tokens = []
|
| 109 |
logprobs = []
|
| 110 |
-
top_alternatives = []
|
| 111 |
for entry in content:
|
| 112 |
if not isinstance(entry, dict):
|
|
|
|
| 113 |
continue
|
| 114 |
logprob = ensure_float(entry.get("logprob", None))
|
| 115 |
-
if logprob >= -100000:
|
| 116 |
tokens.append(get_token(entry))
|
| 117 |
logprobs.append(logprob)
|
| 118 |
top_probs = entry.get("top_logprobs", {}) or {}
|
| 119 |
-
finite_top_probs = [
|
| 120 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
if not logprobs or not tokens:
|
| 123 |
-
return (create_empty_figure("Log Probabilities"), None, "No tokens to display.", create_empty_figure("Top Token Log Probabilities"), create_empty_figure("Probability Drops"), 1, 0)
|
| 124 |
|
| 125 |
total_chunks = max(1, (len(logprobs) + chunk_size - 1) // chunk_size)
|
| 126 |
start_idx = chunk * chunk_size
|
| 127 |
end_idx = min((chunk + 1) * chunk_size, len(logprobs))
|
| 128 |
paginated_tokens = tokens[start_idx:end_idx]
|
| 129 |
paginated_logprobs = logprobs[start_idx:end_idx]
|
| 130 |
-
paginated_alternatives = top_alternatives[start_idx:end_idx]
|
| 131 |
|
| 132 |
-
# Main Log Probability Plot
|
| 133 |
main_fig = go.Figure()
|
| 134 |
main_fig.add_trace(go.Scatter(x=list(range(len(paginated_logprobs))), y=paginated_logprobs, mode='markers+lines', name='Log Prob', marker=dict(color='blue')))
|
| 135 |
-
main_fig.update_layout(
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 141 |
drops = [paginated_logprobs[i+1] - paginated_logprobs[i] for i in range(len(paginated_logprobs)-1)]
|
|
|
|
| 142 |
drops_fig.add_trace(go.Bar(x=list(range(len(drops))), y=drops, name='Drop', marker_color='red'))
|
| 143 |
-
drops_fig.update_layout(
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 147 |
max_alternatives = max(len(alts) for alts in paginated_alternatives) if paginated_alternatives else 0
|
| 148 |
-
|
| 149 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 150 |
|
| 151 |
-
|
| 152 |
-
min_prob, max_prob = min(paginated_logprobs), max(paginated_logprobs)
|
| 153 |
-
normalized_probs = [0.5] * len(paginated_logprobs) if max_prob == min_prob else [(lp - min_prob) / (max_prob - min_prob) for lp in paginated_logprobs]
|
| 154 |
-
colored_text = "".join(f'<span style="color: rgb({int(255*(1-p))}, {int(255*p)}, 0);">{tok}</span> ' for tok, p in zip(paginated_tokens, normalized_probs))
|
| 155 |
-
colored_text_html = f"<p>{colored_text.rstrip()}</p>"
|
| 156 |
|
| 157 |
-
#
|
| 158 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 159 |
if paginated_alternatives:
|
| 160 |
-
for i, (
|
| 161 |
-
for alt_tok, prob in
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 165 |
|
| 166 |
-
return (main_fig, df, colored_text_html, alt_fig, drops_fig, total_chunks, chunk)
|
| 167 |
except Exception as e:
|
| 168 |
logger.error("Visualization failed: %s", str(e))
|
| 169 |
-
return (create_empty_figure("Log Probabilities"), None, f"Error: {e}", create_empty_figure("Top Token Log Probabilities"), create_empty_figure("Probability Drops"), 1, 0)
|
| 170 |
-
|
| 171 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 172 |
def analyze_full_trace(json_input):
|
| 173 |
try:
|
| 174 |
data = parse_input(json_input)
|
| 175 |
content = data.get("content", []) if isinstance(data, dict) else data
|
| 176 |
if not isinstance(content, list):
|
| 177 |
-
raise ValueError("Content must be a list")
|
| 178 |
-
|
| 179 |
-
tokens = [get_token(entry) for entry in content if isinstance(entry, dict) and ensure_float(entry.get("logprob", None)) >= -100000]
|
| 180 |
-
logprobs = [[(key, ensure_float(value)) for key, value in (entry.get("top_logprobs", {}) or {}).items() if ensure_float(value) is not None and math.isfinite(ensure_float(value))] for entry in content if isinstance(entry, dict) and ensure_float(entry.get("logprob", None)) >= -100000]
|
| 181 |
|
| 182 |
-
|
| 183 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 184 |
|
| 185 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 186 |
return analysis_html, None, None, None, None, None
|
| 187 |
except Exception as e:
|
| 188 |
logger.error("Trace analysis failed: %s", str(e))
|
| 189 |
return f"Error: {e}", None, None, None, None, None
|
| 190 |
|
| 191 |
-
# Gradio interface
|
| 192 |
try:
|
| 193 |
with gr.Blocks(title="Log Probability Visualizer") as app:
|
| 194 |
gr.Markdown("# Log Probability Visualizer")
|
| 195 |
-
gr.Markdown("Paste your JSON log prob data below to analyze reasoning traces or visualize tokens in chunks of 100.")
|
| 196 |
|
| 197 |
with gr.Tabs():
|
| 198 |
with gr.Tab("Trace Analysis"):
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 202 |
|
| 203 |
with gr.Tab("Visualization"):
|
| 204 |
with gr.Row():
|
| 205 |
-
json_input_viz = gr.Textbox(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 206 |
chunk = gr.Number(value=0, label="Current Chunk", precision=0, minimum=0)
|
|
|
|
| 207 |
with gr.Row():
|
| 208 |
-
plot_output = gr.Plot(label="Log Probability Plot")
|
| 209 |
-
drops_output = gr.Plot(label="Probability Drops")
|
|
|
|
| 210 |
with gr.Row():
|
| 211 |
-
table_output = gr.Dataframe(label="Token Log Probabilities")
|
| 212 |
-
alt_viz_output = gr.Plot(label="Top Token Log Probabilities")
|
|
|
|
| 213 |
with gr.Row():
|
| 214 |
-
text_output = gr.HTML(label="Colored Text")
|
|
|
|
| 215 |
with gr.Row():
|
| 216 |
prev_btn = gr.Button("Previous Chunk")
|
| 217 |
next_btn = gr.Button("Next Chunk")
|
| 218 |
total_chunks_output = gr.Number(label="Total Chunks", interactive=False)
|
| 219 |
|
|
|
|
| 220 |
precomputed_next = gr.State(value=None)
|
| 221 |
|
| 222 |
-
gr.Button("Visualize")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 223 |
|
| 224 |
def update_chunk(json_input, current_chunk, action, precomputed_next=None):
|
| 225 |
-
total_chunks = visualize_logprobs(json_input, 0)[5]
|
| 226 |
if action == "prev" and current_chunk > 0:
|
| 227 |
current_chunk -= 1
|
| 228 |
elif action == "next" and current_chunk < total_chunks - 1:
|
| 229 |
current_chunk += 1
|
|
|
|
|
|
|
|
|
|
| 230 |
return visualize_logprobs(json_input, current_chunk)
|
| 231 |
|
| 232 |
-
prev_btn.click(
|
| 233 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 234 |
|
| 235 |
def trigger_precomputation(json_input, current_chunk):
|
| 236 |
-
|
|
|
|
|
|
|
|
|
|
| 237 |
return gr.update(value=current_chunk)
|
| 238 |
|
| 239 |
-
chunk.change(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 240 |
|
| 241 |
except Exception as e:
|
| 242 |
logger.error("Application startup failed: %s", str(e))
|
|
|
|
| 23 |
logger.debug("Successfully parsed as JSON")
|
| 24 |
return data
|
| 25 |
except json.JSONDecodeError as e:
|
| 26 |
+
logger.error("JSON parsing failed: %s (Input: %s)", str(e), json_input[:100] + "..." if len(json_input) > 100 else json_input)
|
| 27 |
+
raise ValueError(f"Malformed JSON: {str(e)}. Use double quotes for property names (e.g., \"content\") and ensure valid JSON format.")
|
| 28 |
|
| 29 |
# Function to ensure a value is a float
|
| 30 |
def ensure_float(value):
|
| 31 |
if value is None:
|
| 32 |
+
logger.debug("Replacing None logprob with 0.0")
|
| 33 |
+
return 0.0 # Default to 0.0 for None to ensure visualization
|
|
|
|
| 34 |
if isinstance(value, str):
|
| 35 |
try:
|
| 36 |
return float(value)
|
| 37 |
except ValueError:
|
| 38 |
+
logger.error("Failed to convert string '%s' to float", value)
|
| 39 |
+
return 0.0 # Default to 0.0 for invalid strings
|
| 40 |
+
if isinstance(value, (int, float)):
|
| 41 |
+
return float(value)
|
| 42 |
+
return 0.0 # Default for any other type
|
| 43 |
|
| 44 |
+
# Function to get or generate a token value (default to "Unknown" if missing)
|
| 45 |
def get_token(entry):
|
| 46 |
+
token = entry.get("token", "Unknown")
|
| 47 |
+
if token == "Unknown":
|
| 48 |
+
logger.warning("Missing 'token' key for entry: %s, using 'Unknown'", entry)
|
| 49 |
+
return token
|
| 50 |
|
| 51 |
# Function to create an empty Plotly figure
|
| 52 |
def create_empty_figure(title):
|
| 53 |
return go.Figure().update_layout(title=title, xaxis_title="", yaxis_title="", showlegend=False)
|
| 54 |
|
| 55 |
+
# Precompute the next chunk asynchronously
|
| 56 |
async def precompute_chunk(json_input, chunk_size, current_chunk):
|
| 57 |
try:
|
| 58 |
data = parse_input(json_input)
|
| 59 |
content = data.get("content", []) if isinstance(data, dict) else data
|
| 60 |
if not isinstance(content, list):
|
| 61 |
+
raise ValueError("Content must be a list of entries")
|
| 62 |
|
| 63 |
tokens = []
|
| 64 |
logprobs = []
|
| 65 |
top_alternatives = []
|
| 66 |
for entry in content:
|
| 67 |
if not isinstance(entry, dict):
|
| 68 |
+
logger.warning("Skipping non-dictionary entry: %s", entry)
|
| 69 |
continue
|
| 70 |
logprob = ensure_float(entry.get("logprob", None))
|
| 71 |
+
if logprob >= -100000: # Include all entries with default 0.0
|
| 72 |
tokens.append(get_token(entry))
|
| 73 |
logprobs.append(logprob)
|
| 74 |
+
top_probs = entry.get("top_logprobs", {})
|
| 75 |
+
if top_probs is None:
|
| 76 |
+
logger.debug("top_logprobs is None for token: %s, using empty dict", get_token(entry))
|
| 77 |
+
top_probs = {}
|
| 78 |
+
finite_top_probs = []
|
| 79 |
+
for key, value in top_probs.items():
|
| 80 |
+
float_value = ensure_float(value)
|
| 81 |
+
if float_value is not None and math.isfinite(float_value):
|
| 82 |
+
finite_top_probs.append((key, float_value))
|
| 83 |
+
sorted_probs = sorted(finite_top_probs, key=lambda x: x[1], reverse=True)
|
| 84 |
+
top_alternatives.append(sorted_probs)
|
| 85 |
|
| 86 |
if not tokens or not logprobs:
|
| 87 |
return None, None, None
|
|
|
|
| 92 |
if start_idx >= len(tokens):
|
| 93 |
return None, None, None
|
| 94 |
|
| 95 |
+
paginated_tokens = tokens[start_idx:end_idx]
|
| 96 |
+
paginated_logprobs = logprobs[start_idx:end_idx]
|
| 97 |
+
paginated_alternatives = top_alternatives[start_idx:end_idx]
|
| 98 |
+
|
| 99 |
+
return paginated_tokens, paginated_logprobs, paginated_alternatives
|
| 100 |
except Exception as e:
|
| 101 |
logger.error("Precomputation failed for chunk %d: %s", current_chunk + 1, str(e))
|
| 102 |
return None, None, None
|
|
|
|
| 114 |
loop.close()
|
| 115 |
return result
|
| 116 |
|
| 117 |
+
# Function to process and visualize a chunk of log probs with dynamic top_logprobs
|
| 118 |
def visualize_logprobs(json_input, chunk=0, chunk_size=100):
|
| 119 |
try:
|
| 120 |
data = parse_input(json_input)
|
| 121 |
content = data.get("content", []) if isinstance(data, dict) else data
|
| 122 |
if not isinstance(content, list):
|
| 123 |
+
raise ValueError("Content must be a list of entries")
|
| 124 |
|
| 125 |
tokens = []
|
| 126 |
logprobs = []
|
| 127 |
+
top_alternatives = [] # List to store all top_logprobs (dynamic length)
|
| 128 |
for entry in content:
|
| 129 |
if not isinstance(entry, dict):
|
| 130 |
+
logger.warning("Skipping non-dictionary entry: %s", entry)
|
| 131 |
continue
|
| 132 |
logprob = ensure_float(entry.get("logprob", None))
|
| 133 |
+
if logprob >= -100000: # Include all entries with default 0.0
|
| 134 |
tokens.append(get_token(entry))
|
| 135 |
logprobs.append(logprob)
|
| 136 |
top_probs = entry.get("top_logprobs", {}) or {}
|
| 137 |
+
finite_top_probs = []
|
| 138 |
+
for key, value in top_probs.items():
|
| 139 |
+
float_value = ensure_float(value)
|
| 140 |
+
if float_value is not None and math.isfinite(float_value):
|
| 141 |
+
finite_top_probs.append((key, float_value))
|
| 142 |
+
sorted_probs = sorted(finite_top_probs, key=lambda x: x[1], reverse=True)
|
| 143 |
+
top_alternatives.append(sorted_probs)
|
| 144 |
|
| 145 |
if not logprobs or not tokens:
|
| 146 |
+
return (create_empty_figure("Log Probabilities of Generated Tokens"), None, "No tokens to display.", create_empty_figure("Top Token Log Probabilities"), create_empty_figure("Significant Probability Drops"), 1, 0)
|
| 147 |
|
| 148 |
total_chunks = max(1, (len(logprobs) + chunk_size - 1) // chunk_size)
|
| 149 |
start_idx = chunk * chunk_size
|
| 150 |
end_idx = min((chunk + 1) * chunk_size, len(logprobs))
|
| 151 |
paginated_tokens = tokens[start_idx:end_idx]
|
| 152 |
paginated_logprobs = logprobs[start_idx:end_idx]
|
| 153 |
+
paginated_alternatives = top_alternatives[start_idx:end_idx] if top_alternatives else []
|
| 154 |
|
| 155 |
+
# Main Log Probability Plot (Interactive Plotly)
|
| 156 |
main_fig = go.Figure()
|
| 157 |
main_fig.add_trace(go.Scatter(x=list(range(len(paginated_logprobs))), y=paginated_logprobs, mode='markers+lines', name='Log Prob', marker=dict(color='blue')))
|
| 158 |
+
main_fig.update_layout(
|
| 159 |
+
title=f"Log Probabilities of Generated Tokens (Chunk {chunk + 1})",
|
| 160 |
+
xaxis_title="Token Position (within chunk)",
|
| 161 |
+
yaxis_title="Log Probability",
|
| 162 |
+
hovermode="closest",
|
| 163 |
+
clickmode='event+select'
|
| 164 |
+
)
|
| 165 |
+
main_fig.update_traces(
|
| 166 |
+
customdata=[f"Token: {tok}, Log Prob: {prob:.4f}, Position: {i+start_idx}" for i, (tok, prob) in enumerate(zip(paginated_tokens, paginated_logprobs))],
|
| 167 |
+
hovertemplate='<b>%{customdata}</b><extra></extra>'
|
| 168 |
+
)
|
| 169 |
+
|
| 170 |
+
# Probability Drop Analysis (Interactive Plotly)
|
| 171 |
+
if len(paginated_logprobs) < 2:
|
| 172 |
+
drops_fig = create_empty_figure(f"Significant Probability Drops (Chunk {chunk + 1})")
|
| 173 |
+
else:
|
| 174 |
drops = [paginated_logprobs[i+1] - paginated_logprobs[i] for i in range(len(paginated_logprobs)-1)]
|
| 175 |
+
drops_fig = go.Figure()
|
| 176 |
drops_fig.add_trace(go.Bar(x=list(range(len(drops))), y=drops, name='Drop', marker_color='red'))
|
| 177 |
+
drops_fig.update_layout(
|
| 178 |
+
title=f"Significant Probability Drops (Chunk {chunk + 1})",
|
| 179 |
+
xaxis_title="Token Position (within chunk)",
|
| 180 |
+
yaxis_title="Log Probability Drop",
|
| 181 |
+
hovermode="closest",
|
| 182 |
+
clickmode='event+select'
|
| 183 |
+
)
|
| 184 |
+
drops_fig.update_traces(
|
| 185 |
+
customdata=[f"Drop: {drop:.4f}, From: {paginated_tokens[i]} to {paginated_tokens[i+1]}, Position: {i+start_idx}" for i, drop in enumerate(drops)],
|
| 186 |
+
hovertemplate='<b>%{customdata}</b><extra></extra>'
|
| 187 |
+
)
|
| 188 |
+
|
| 189 |
+
# Create DataFrame for the table with dynamic top_logprobs
|
| 190 |
+
table_data = []
|
| 191 |
max_alternatives = max(len(alts) for alts in paginated_alternatives) if paginated_alternatives else 0
|
| 192 |
+
for i, entry in enumerate(content[start_idx:end_idx]):
|
| 193 |
+
if not isinstance(entry, dict):
|
| 194 |
+
continue
|
| 195 |
+
logprob = ensure_float(entry.get("logprob", None))
|
| 196 |
+
if logprob >= -100000 and "top_logprobs" in entry:
|
| 197 |
+
token = get_token(entry)
|
| 198 |
+
top_logprobs = entry.get("top_logprobs", {}) or {}
|
| 199 |
+
finite_top_probs = []
|
| 200 |
+
for key, value in top_logprobs.items():
|
| 201 |
+
float_value = ensure_float(value)
|
| 202 |
+
if float_value is not None and math.isfinite(float_value):
|
| 203 |
+
finite_top_probs.append((key, float_value))
|
| 204 |
+
sorted_probs = sorted(finite_top_probs, key=lambda x: x[1], reverse=True)
|
| 205 |
+
row = [token, f"{logprob:.4f}"]
|
| 206 |
+
for alt_token, alt_logprob in sorted_probs[:max_alternatives]:
|
| 207 |
+
row.append(f"{alt_token}: {alt_logprob:.4f}")
|
| 208 |
+
while len(row) < 2 + max_alternatives:
|
| 209 |
+
row.append("")
|
| 210 |
+
table_data.append(row)
|
| 211 |
|
| 212 |
+
df = pd.DataFrame(table_data, columns=["Token", "Log Prob"] + [f"Alt {i+1}" for i in range(max_alternatives)]) if table_data else None
|
|
|
|
|
|
|
|
|
|
|
|
|
| 213 |
|
| 214 |
+
# Generate colored text (for the current chunk)
|
| 215 |
+
if paginated_logprobs:
|
| 216 |
+
min_logprob = min(paginated_logprobs)
|
| 217 |
+
max_logprob = max(paginated_logprobs)
|
| 218 |
+
normalized_probs = [0.5] * len(paginated_logprobs) if max_logprob == min_logprob else \
|
| 219 |
+
[(lp - min_logprob) / (max_logprob - min_logprob) for lp in paginated_logprobs]
|
| 220 |
+
|
| 221 |
+
colored_text = ""
|
| 222 |
+
for i, (token, norm_prob) in enumerate(zip(paginated_tokens, normalized_probs)):
|
| 223 |
+
r = int(255 * (1 - norm_prob)) # Red for low confidence
|
| 224 |
+
g = int(255 * norm_prob) # Green for high confidence
|
| 225 |
+
b = 0
|
| 226 |
+
color = f"rgb({r}, {g}, {b})"
|
| 227 |
+
colored_text += f'<span style="color: {color}; font-weight: bold;">{token}</span>'
|
| 228 |
+
if i < len(paginated_tokens) - 1:
|
| 229 |
+
colored_text += " "
|
| 230 |
+
colored_text_html = f"<p>{colored_text}</p>"
|
| 231 |
+
else:
|
| 232 |
+
colored_text_html = "No tokens to display in this chunk."
|
| 233 |
+
|
| 234 |
+
# Top Token Log Probabilities (Interactive Plotly, dynamic length, for the current chunk)
|
| 235 |
+
alt_viz_fig = create_empty_figure(f"Top Token Log Probabilities (Chunk {chunk + 1})") if not paginated_alternatives else go.Figure()
|
| 236 |
if paginated_alternatives:
|
| 237 |
+
for i, (token, probs) in enumerate(zip(paginated_tokens, paginated_alternatives)):
|
| 238 |
+
for j, (alt_tok, prob) in enumerate(probs):
|
| 239 |
+
alt_viz_fig.add_trace(go.Bar(x=[f"{token} (Pos {i+start_idx})"], y=[prob], name=f"{alt_tok}", marker_color=['blue', 'green', 'red', 'purple', 'orange'][:len(probs)]))
|
| 240 |
+
alt_viz_fig.update_layout(
|
| 241 |
+
title=f"Top Token Log Probabilities (Chunk {chunk + 1})",
|
| 242 |
+
xaxis_title="Token (Position)",
|
| 243 |
+
yaxis_title="Log Probability",
|
| 244 |
+
barmode='stack',
|
| 245 |
+
hovermode="closest",
|
| 246 |
+
clickmode='event+select'
|
| 247 |
+
)
|
| 248 |
+
alt_viz_fig.update_traces(
|
| 249 |
+
customdata=[f"Token: {tok}, Alt: {alt}, Log Prob: {prob:.4f}, Position: {i+start_idx}" for i, (tok, alts) in enumerate(zip(paginated_tokens, paginated_alternatives)) for alt, prob in alts],
|
| 250 |
+
hovertemplate='<b>%{customdata}</b><extra></extra>'
|
| 251 |
+
)
|
| 252 |
+
|
| 253 |
+
return (main_fig, df, colored_text_html, alt_viz_fig, drops_fig, total_chunks, chunk)
|
| 254 |
|
|
|
|
| 255 |
except Exception as e:
|
| 256 |
logger.error("Visualization failed: %s", str(e))
|
| 257 |
+
return (create_empty_figure("Log Probabilities of Generated Tokens"), None, f"Error: {e}", create_empty_figure("Top Token Log Probabilities"), create_empty_figure("Significant Probability Drops"), 1, 0)
|
| 258 |
+
|
| 259 |
+
# Analysis functions for detecting correct vs. incorrect traces
|
| 260 |
+
def analyze_confidence_signature(logprobs, tokens):
|
| 261 |
+
if not logprobs or not tokens:
|
| 262 |
+
return "No data for confidence signature analysis.", None
|
| 263 |
+
top_probs = [lps[0][1] if lps and lps[0][1] is not None else -float('inf') for lps in logprobs] # Handle empty or None
|
| 264 |
+
if not any(p != -float('inf') for p in top_probs):
|
| 265 |
+
return "No valid log probabilities for confidence analysis.", None
|
| 266 |
+
moving_avg = np.convolve(top_probs, np.ones(20) / 20, mode='valid') # 20-token window
|
| 267 |
+
drops = np.where(np.diff(moving_avg) < -0.15)[0] # Significant drops
|
| 268 |
+
if not drops.size:
|
| 269 |
+
return "No significant confidence drops detected.", None
|
| 270 |
+
drop_positions = [(i, tokens[i + 19] if i + 19 < len(tokens) else "End of trace") for i in drops]
|
| 271 |
+
return "Significant confidence drops detected at positions:", drop_positions
|
| 272 |
+
|
| 273 |
+
def detect_interpretation_pivots(logprobs, tokens):
|
| 274 |
+
if not logprobs or not tokens:
|
| 275 |
+
return "No data for interpretation pivot detection.", None
|
| 276 |
+
pivots = []
|
| 277 |
+
reconsideration_tokens = ["wait", "but", "actually", "however", "hmm"]
|
| 278 |
+
for i, (token, lps) in enumerate(zip(tokens, logprobs)):
|
| 279 |
+
if not lps:
|
| 280 |
+
continue
|
| 281 |
+
for rt in reconsideration_tokens:
|
| 282 |
+
for t, p in lps:
|
| 283 |
+
if t.lower() == rt and p > -2.5: # High probability
|
| 284 |
+
context = tokens[max(0, i-50):i]
|
| 285 |
+
pivots.append((i, rt, context))
|
| 286 |
+
if not pivots:
|
| 287 |
+
return "No interpretation pivots detected.", None
|
| 288 |
+
return "Interpretation pivots detected:", pivots
|
| 289 |
+
|
| 290 |
+
def calculate_decision_entropy(logprobs):
|
| 291 |
+
if not logprobs:
|
| 292 |
+
return "No data for entropy spike detection.", None
|
| 293 |
+
entropies = []
|
| 294 |
+
for lps in logprobs:
|
| 295 |
+
if not lps:
|
| 296 |
+
entropies.append(0.0)
|
| 297 |
+
continue
|
| 298 |
+
probs = [math.exp(p) for _, p in lps if p is not None] # Convert log probs to probabilities, handle None
|
| 299 |
+
if not probs or sum(probs) == 0:
|
| 300 |
+
entropies.append(0.0)
|
| 301 |
+
continue
|
| 302 |
+
entropy = -sum(p * math.log(p) for p in probs if p > 0)
|
| 303 |
+
entropies.append(entropy)
|
| 304 |
+
baseline = np.percentile(entropies, 75) if entropies else 0.0
|
| 305 |
+
spikes = [i for i, e in enumerate(entropies) if e > baseline * 1.5 and baseline > 0]
|
| 306 |
+
if not spikes:
|
| 307 |
+
return "No entropy spikes detected at decision points.", None
|
| 308 |
+
return "Entropy spikes detected at positions:", spikes
|
| 309 |
+
|
| 310 |
+
def analyze_conclusion_competition(logprobs, tokens):
|
| 311 |
+
if not logprobs or not tokens:
|
| 312 |
+
return "No data for conclusion competition analysis.", None
|
| 313 |
+
conclusion_indices = [i for i, t in enumerate(tokens) if any(marker in t.lower() for marker in ["therefore", "thus", "boxed", "answer"])]
|
| 314 |
+
if not conclusion_indices:
|
| 315 |
+
return "No conclusion markers found in trace.", None
|
| 316 |
+
gaps = []
|
| 317 |
+
conclusion_idx = conclusion_indices[-1]
|
| 318 |
+
end_range = min(conclusion_idx + 50, len(logprobs))
|
| 319 |
+
for idx in range(conclusion_idx, end_range):
|
| 320 |
+
if idx < len(logprobs) and len(logprobs[idx]) >= 2 and logprobs[idx][0][1] is not None and logprobs[idx][1][1] is not None:
|
| 321 |
+
gap = logprobs[idx][0][1] - logprobs[idx][1][1]
|
| 322 |
+
gaps.append(gap)
|
| 323 |
+
if not gaps:
|
| 324 |
+
return "No conclusion competition data available.", None
|
| 325 |
+
mean_gap = np.mean(gaps)
|
| 326 |
+
return f"Mean probability gap at conclusion: {mean_gap:.4f} (higher indicates more confident conclusion)", None
|
| 327 |
+
|
| 328 |
+
def analyze_verification_signals(logprobs, tokens):
|
| 329 |
+
if not logprobs or not tokens:
|
| 330 |
+
return "No data for verification signal analysis.", None
|
| 331 |
+
verification_terms = ["verify", "check", "confirm", "ensure", "double"]
|
| 332 |
+
verification_probs = []
|
| 333 |
+
for lps in logprobs:
|
| 334 |
+
if not lps:
|
| 335 |
+
continue
|
| 336 |
+
max_v_prob = -float('inf')
|
| 337 |
+
for token, prob in lps:
|
| 338 |
+
if any(v_term in token.lower() for v_term in verification_terms) and prob is not None:
|
| 339 |
+
max_v_prob = max(max_v_prob, prob)
|
| 340 |
+
if max_v_prob > -float('inf'):
|
| 341 |
+
verification_probs.append(max_v_prob)
|
| 342 |
+
if not verification_probs:
|
| 343 |
+
return "No verification signals detected.", None
|
| 344 |
+
count, mean_prob = len(verification_probs), np.mean(verification_probs)
|
| 345 |
+
return f"Verification signals found: {count} instances, mean probability: {mean_prob:.4f}", None
|
| 346 |
+
|
| 347 |
+
def detect_semantic_inversions(logprobs, tokens):
|
| 348 |
+
if not logprobs or not tokens:
|
| 349 |
+
return "No data for semantic inversion detection.", None
|
| 350 |
+
inversion_pairs = [("more", "less"), ("larger", "smaller"), ("winning", "losing"), ("increase", "decrease"), ("greater", "lesser"), ("positive", "negative")]
|
| 351 |
+
inversions = []
|
| 352 |
+
for i, (token, lps) in enumerate(zip(tokens, logprobs)):
|
| 353 |
+
if not lps:
|
| 354 |
+
continue
|
| 355 |
+
for pos, neg in inversion_pairs:
|
| 356 |
+
if token.lower() == pos:
|
| 357 |
+
for t, p in lps:
|
| 358 |
+
if t.lower() == neg and p > -3.0 and p is not None:
|
| 359 |
+
inversions.append((i, pos, neg, p))
|
| 360 |
+
elif token.lower() == neg:
|
| 361 |
+
for t, p in lps:
|
| 362 |
+
if t.lower() == pos and p > -3.0 and p is not None:
|
| 363 |
+
inversions.append((i, neg, pos, p))
|
| 364 |
+
if not inversions:
|
| 365 |
+
return "No semantic inversions detected.", None
|
| 366 |
+
return "Semantic inversions detected:", inversions
|
| 367 |
+
|
| 368 |
+
# Function to perform full trace analysis
|
| 369 |
def analyze_full_trace(json_input):
|
| 370 |
try:
|
| 371 |
data = parse_input(json_input)
|
| 372 |
content = data.get("content", []) if isinstance(data, dict) else data
|
| 373 |
if not isinstance(content, list):
|
| 374 |
+
raise ValueError("Content must be a list of entries")
|
|
|
|
|
|
|
|
|
|
| 375 |
|
| 376 |
+
tokens = []
|
| 377 |
+
logprobs = []
|
| 378 |
+
for entry in content:
|
| 379 |
+
if not isinstance(entry, dict):
|
| 380 |
+
logger.warning("Skipping non-dictionary entry: %s", entry)
|
| 381 |
+
continue
|
| 382 |
+
logprob = ensure_float(entry.get("logprob", None))
|
| 383 |
+
if logprob >= -100000:
|
| 384 |
+
tokens.append(get_token(entry))
|
| 385 |
+
top_probs = entry.get("top_logprobs", {}) or {}
|
| 386 |
+
finite_top_probs = [(key, ensure_float(value)) for key, value in top_probs.items() if ensure_float(value) is not None and math.isfinite(ensure_float(value))]
|
| 387 |
+
logprobs.append(finite_top_probs)
|
| 388 |
|
| 389 |
+
if not logprobs or not tokens:
|
| 390 |
+
return "No valid data for trace analysis.", None, None, None, None, None
|
| 391 |
+
|
| 392 |
+
confidence_result, confidence_data = analyze_confidence_signature(logprobs, tokens)
|
| 393 |
+
pivot_result, pivot_data = detect_interpretation_pivots(logprobs, tokens)
|
| 394 |
+
entropy_result, entropy_data = calculate_decision_entropy(logprobs)
|
| 395 |
+
conclusion_result, conclusion_data = analyze_conclusion_competition(logprobs, tokens)
|
| 396 |
+
verification_result, verification_data = analyze_verification_signals(logprobs, tokens)
|
| 397 |
+
inversion_result, inversion_data = detect_semantic_inversions(logprobs, tokens)
|
| 398 |
+
|
| 399 |
+
analysis_html = f"""
|
| 400 |
+
<h3>Trace Analysis Results</h3>
|
| 401 |
+
<ul>
|
| 402 |
+
<li><strong>Confidence Signature:</strong> {confidence_result}</li>
|
| 403 |
+
{f"<ul><li>Positions: {', '.join(str(pos) for pos, tok in confidence_data)}</li></ul>" if confidence_data else ""}
|
| 404 |
+
<li><strong>Interpretation Pivots:</strong> {pivot_result}</li>
|
| 405 |
+
{f"<ul><li>Positions: {', '.join(str(pos) for pos, _, _ in pivot_data)}</li></ul>" if pivot_data else ""}
|
| 406 |
+
<li><strong>Decision Entropy Spikes:</strong> {entropy_result}</li>
|
| 407 |
+
{f"<ul><li>Positions: {', '.join(str(pos) for pos in entropy_data)}</li></ul>" if entropy_data else ""}
|
| 408 |
+
<li><strong>Conclusion Competition:</strong> {conclusion_result}</li>
|
| 409 |
+
<li><strong>Verification Signals:</strong> {verification_result}</li>
|
| 410 |
+
<li><strong>Semantic Inversions:</strong> {inversion_result}</li>
|
| 411 |
+
{f"<ul><li>Positions: {', '.join(str(pos) for pos, _, _, _ in inversion_data)}</li></ul>" if inversion_data else ""}
|
| 412 |
+
</ul>
|
| 413 |
+
"""
|
| 414 |
return analysis_html, None, None, None, None, None
|
| 415 |
except Exception as e:
|
| 416 |
logger.error("Trace analysis failed: %s", str(e))
|
| 417 |
return f"Error: {e}", None, None, None, None, None
|
| 418 |
|
| 419 |
+
# Gradio interface with two tabs
|
| 420 |
try:
|
| 421 |
with gr.Blocks(title="Log Probability Visualizer") as app:
|
| 422 |
gr.Markdown("# Log Probability Visualizer")
|
| 423 |
+
gr.Markdown("Paste your JSON log prob data below to analyze reasoning traces or visualize tokens in chunks of 100. Fixed filter ≥ -100000, dynamic number of top_logprobs, handles missing or null fields. Next chunk is precomputed proactively.")
|
| 424 |
|
| 425 |
with gr.Tabs():
|
| 426 |
with gr.Tab("Trace Analysis"):
|
| 427 |
+
with gr.Row():
|
| 428 |
+
json_input_analysis = gr.Textbox(
|
| 429 |
+
label="JSON Input for Trace Analysis",
|
| 430 |
+
lines=10,
|
| 431 |
+
placeholder='{"content": [{"bytes": [44], "logprob": 0.0, "token": ",", "top_logprobs": {" so": -13.8046875, ".": -13.8046875, ",": -13.640625}}]}'
|
| 432 |
+
)
|
| 433 |
+
with gr.Row():
|
| 434 |
+
analysis_output = gr.HTML(label="Trace Analysis Results")
|
| 435 |
+
|
| 436 |
+
btn_analyze = gr.Button("Analyze Trace")
|
| 437 |
+
btn_analyze.click(
|
| 438 |
+
fn=analyze_full_trace,
|
| 439 |
+
inputs=[json_input_analysis],
|
| 440 |
+
outputs=[analysis_output, gr.State(), gr.State(), gr.State(), gr.State(), gr.State()],
|
| 441 |
+
)
|
| 442 |
|
| 443 |
with gr.Tab("Visualization"):
|
| 444 |
with gr.Row():
|
| 445 |
+
json_input_viz = gr.Textbox(
|
| 446 |
+
label="JSON Input for Visualization",
|
| 447 |
+
lines=10,
|
| 448 |
+
placeholder='{"content": [{"bytes": [44], "logprob": 0.0, "token": ",", "top_logprobs": {" so": -13.8046875, ".": -13.8046875, ",": -13.640625}}]}'
|
| 449 |
+
)
|
| 450 |
chunk = gr.Number(value=0, label="Current Chunk", precision=0, minimum=0)
|
| 451 |
+
|
| 452 |
with gr.Row():
|
| 453 |
+
plot_output = gr.Plot(label="Log Probability Plot (Click for Tokens)")
|
| 454 |
+
drops_output = gr.Plot(label="Probability Drops (Click for Details)")
|
| 455 |
+
|
| 456 |
with gr.Row():
|
| 457 |
+
table_output = gr.Dataframe(label="Token Log Probabilities and Top Alternatives")
|
| 458 |
+
alt_viz_output = gr.Plot(label="Top Token Log Probabilities (Click for Details)")
|
| 459 |
+
|
| 460 |
with gr.Row():
|
| 461 |
+
text_output = gr.HTML(label="Colored Text (Confidence Visualization)")
|
| 462 |
+
|
| 463 |
with gr.Row():
|
| 464 |
prev_btn = gr.Button("Previous Chunk")
|
| 465 |
next_btn = gr.Button("Next Chunk")
|
| 466 |
total_chunks_output = gr.Number(label="Total Chunks", interactive=False)
|
| 467 |
|
| 468 |
+
# Precomputed next chunk state (hidden)
|
| 469 |
precomputed_next = gr.State(value=None)
|
| 470 |
|
| 471 |
+
btn_viz = gr.Button("Visualize")
|
| 472 |
+
btn_viz.click(
|
| 473 |
+
fn=visualize_logprobs,
|
| 474 |
+
inputs=[json_input_viz, chunk],
|
| 475 |
+
outputs=[plot_output, table_output, text_output, alt_viz_output, drops_output, total_chunks_output, chunk],
|
| 476 |
+
)
|
| 477 |
|
| 478 |
def update_chunk(json_input, current_chunk, action, precomputed_next=None):
|
| 479 |
+
total_chunks = visualize_logprobs(json_input, 0)[5] # Get total chunks
|
| 480 |
if action == "prev" and current_chunk > 0:
|
| 481 |
current_chunk -= 1
|
| 482 |
elif action == "next" and current_chunk < total_chunks - 1:
|
| 483 |
current_chunk += 1
|
| 484 |
+
if precomputed_next and all(precomputed_next):
|
| 485 |
+
logger.debug("Using precomputed next chunk for chunk %d", current_chunk)
|
| 486 |
+
return visualize_logprobs(json_input, current_chunk)
|
| 487 |
return visualize_logprobs(json_input, current_chunk)
|
| 488 |
|
| 489 |
+
prev_btn.click(
|
| 490 |
+
fn=update_chunk,
|
| 491 |
+
inputs=[json_input_viz, chunk, gr.State(value="prev"), precomputed_next],
|
| 492 |
+
outputs=[plot_output, table_output, text_output, alt_viz_output, drops_output, total_chunks_output, chunk],
|
| 493 |
+
)
|
| 494 |
+
|
| 495 |
+
next_btn.click(
|
| 496 |
+
fn=update_chunk,
|
| 497 |
+
inputs=[json_input_viz, chunk, gr.State(value="next"), precomputed_next],
|
| 498 |
+
outputs=[plot_output, table_output, text_output, alt_viz_output, drops_output, total_chunks_output, chunk],
|
| 499 |
+
)
|
| 500 |
|
| 501 |
def trigger_precomputation(json_input, current_chunk):
|
| 502 |
+
try:
|
| 503 |
+
threading.Thread(target=precompute_next_chunk_sync, args=(json_input, current_chunk), daemon=True).start()
|
| 504 |
+
except Exception as e:
|
| 505 |
+
logger.error("Precomputation trigger failed: %s", str(e))
|
| 506 |
return gr.update(value=current_chunk)
|
| 507 |
|
| 508 |
+
chunk.change(
|
| 509 |
+
fn=trigger_precomputation,
|
| 510 |
+
inputs=[json_input_viz, chunk],
|
| 511 |
+
outputs=[chunk],
|
| 512 |
+
)
|
| 513 |
|
| 514 |
except Exception as e:
|
| 515 |
logger.error("Application startup failed: %s", str(e))
|