Spaces:
Sleeping
Sleeping
File size: 11,209 Bytes
6ae9b63 7b90f10 6ae9b63 1b8513e 6ae9b63 1b8513e 6ae9b63 1b8513e 6ae9b63 1b8513e 6ae9b63 1b8513e 6ae9b63 1b8513e 6ae9b63 7b90f10 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 | """Token-level activation heatmaps, feature dashboards, and visualization utilities.
Generates interactive Plotly visualizations for the dashboard:
- Token-level feature activation heatmaps
- Feature activation distributions
- Steered vs. unsteered comparison displays
- Layer-wise feature activity plots
"""
import numpy as np
import plotly.graph_objects as go
import plotly.express as px
from plotly.subplots import make_subplots
from typing import Optional
def create_token_heatmap(
str_tokens: list[str],
activations: list[float],
feature_idx: int,
description: str = "",
colorscale: str = "YlOrRd",
) -> go.Figure:
"""Create a heatmap showing feature activation per token.
Displays tokens along the x-axis with color intensity proportional
to the feature's activation on that token.
"""
# Clean up token strings for display
display_tokens = [t.replace("▁", " ").replace("Ġ", " ") for t in str_tokens]
# Reshape activations for heatmap (1 x n_tokens)
z = np.array(activations).reshape(1, -1)
fig = go.Figure(
data=go.Heatmap(
z=z,
x=display_tokens,
y=["Activation"],
colorscale=colorscale,
text=[[f"{v:.3f}" for v in activations]],
texttemplate="%{text}",
textfont={"size": 10},
hovertemplate="Token: %{x}<br>Activation: %{z:.4f}<extra></extra>",
)
)
title = f"Feature #{feature_idx}"
if description:
title += f": {description[:80]}"
fig.update_layout(
title=title,
xaxis_title="Token",
height=150,
margin=dict(l=60, r=20, t=40, b=40),
xaxis=dict(tickangle=45),
)
return fig
def create_multi_feature_heatmap(
str_tokens: list[str],
feature_data: list[dict],
max_features: int = 10,
colorscale: str = "YlOrRd",
) -> go.Figure:
"""Create a heatmap showing multiple features' activations across tokens.
Each row is a feature, each column is a token. Color intensity shows
activation strength.
"""
display_tokens = [t.replace("▁", " ").replace("Ġ", " ") for t in str_tokens]
data = feature_data[:max_features]
n_features = len(data)
# Build the z-matrix: [n_features x n_tokens]
z = np.zeros((n_features, len(str_tokens)))
y_labels = []
for i, feat in enumerate(data):
acts = feat["per_token_activations"]
z[i, : len(acts)] = acts
desc = feat["description"][:40]
y_labels.append(f"#{feat['feature_idx']}: {desc}")
fig = go.Figure(
data=go.Heatmap(
z=z,
x=display_tokens,
y=y_labels,
colorscale=colorscale,
hovertemplate="Token: %{x}<br>Feature: %{y}<br>Activation: %{z:.4f}<extra></extra>",
)
)
fig.update_layout(
title="Top Active Features by Token (<bos> token skipped)",
xaxis_title="Token",
yaxis_title="Feature",
height=max(300, 60 * n_features),
margin=dict(l=200, r=20, t=40, b=60),
xaxis=dict(tickangle=45),
)
return fig
def create_activation_histogram(
activations: list[float],
feature_idx: int,
description: str = "",
n_bins: int = 50,
) -> go.Figure:
"""Create a histogram of feature activations across tokens."""
acts = np.array(activations)
nonzero = acts[acts > 0]
fig = make_subplots(rows=1, cols=1)
if len(nonzero) > 0:
fig.add_trace(
go.Histogram(
x=nonzero,
nbinsx=n_bins,
name="Non-zero activations",
marker_color="steelblue",
)
)
title = f"Feature #{feature_idx} Activation Distribution"
if description:
title += f"\n{description[:80]}"
sparsity = 1.0 - (len(nonzero) / len(acts)) if len(acts) > 0 else 1.0
fig.update_layout(
title=title,
xaxis_title="Activation Value",
yaxis_title="Count",
height=300,
margin=dict(l=60, r=20, t=60, b=40),
annotations=[
dict(
text=f"Sparsity: {sparsity:.1%} | Active: {len(nonzero)}/{len(acts)}",
xref="paper",
yref="paper",
x=0.95,
y=0.95,
showarrow=False,
font=dict(size=11),
)
],
)
return fig
def create_steering_comparison(
prompt: str,
unsteered: str,
steered: str,
interventions: list[dict],
) -> str:
"""Create an HTML comparison of steered vs. unsteered text.
Returns formatted HTML string for display in Gradio.
"""
import html
prompt_safe = html.escape(prompt)
unsteered_safe = html.escape(unsteered)
steered_safe = html.escape(steered)
intervention_desc = ", ".join(
f"Feature #{i['feature_idx']} (strength={i['strength']:.1f})"
for i in interventions
)
markup = f"""
<div style="font-family: monospace; padding: 10px;">
<h3>Prompt</h3>
<p style="background: #f0f0f0; padding: 10px; border-radius: 5px;">{prompt_safe}</p>
<div style="display: flex; gap: 20px;">
<div style="flex: 1;">
<h3 style="color: #666;">Unsteered</h3>
<p style="background: #f8f8f8; padding: 10px; border-radius: 5px;
border-left: 3px solid #ccc; white-space: pre-wrap;">{unsteered_safe}</p>
</div>
<div style="flex: 1;">
<h3 style="color: #2196F3;">Steered</h3>
<p style="background: #f0f8ff; padding: 10px; border-radius: 5px;
border-left: 3px solid #2196F3; white-space: pre-wrap;">{steered_safe}</p>
</div>
</div>
<p style="color: #888; font-size: 0.9em;">
Interventions: {intervention_desc}
</p>
</div>
"""
return markup
def create_top_predictions_comparison(
clean_tokens: list[dict],
steered_tokens: list[dict],
kl_divergence: float,
) -> go.Figure:
"""Create a side-by-side bar chart comparing top predicted tokens.
Shows how steering changes the model's next-token distribution.
"""
fig = make_subplots(
rows=1,
cols=2,
subplot_titles=["Unsteered Predictions", "Steered Predictions"],
horizontal_spacing=0.15,
)
# Clean predictions
fig.add_trace(
go.Bar(
x=[t["prob"] for t in clean_tokens],
y=[t["token"] for t in clean_tokens],
orientation="h",
marker_color="lightgray",
name="Unsteered",
),
row=1,
col=1,
)
# Steered predictions
fig.add_trace(
go.Bar(
x=[t["prob"] for t in steered_tokens],
y=[t["token"] for t in steered_tokens],
orientation="h",
marker_color="steelblue",
name="Steered",
),
row=1,
col=2,
)
fig.update_layout(
title=f"Next-Token Predictions (KL Divergence: {kl_divergence:.4f})",
height=400,
showlegend=False,
margin=dict(l=80, r=20, t=60, b=40),
)
fig.update_xaxes(title_text="Probability", row=1, col=1)
fig.update_xaxes(title_text="Probability", row=1, col=2)
return fig
def create_layer_activity_plot(
layer_activations: dict[int, float],
feature_idx: int,
description: str = "",
) -> go.Figure:
"""Plot feature activation strength across layers.
Shows which layers a feature is most active in, giving insight
into where in the model's computation the feature matters.
"""
layers = sorted(layer_activations.keys())
values = [layer_activations[l] for l in layers]
fig = go.Figure(
data=go.Bar(
x=[f"Layer {l}" for l in layers],
y=values,
marker_color="steelblue",
)
)
title = f"Feature #{feature_idx} Activity by Layer"
if description:
title += f"\n{description[:60]}"
fig.update_layout(
title=title,
xaxis_title="Layer",
yaxis_title="Mean Activation",
height=350,
margin=dict(l=60, r=20, t=60, b=60),
)
return fig
def create_logit_attribution_chart(
top_positive: list[dict],
top_negative: list[dict],
bias: float,
error: float,
target_token: str,
total_logit: float,
descriptions: Optional[dict[int, str]] = None,
) -> go.Figure:
"""Create a horizontal bar chart of per-feature logit contributions.
Positive contributions shown in blue (right), negative in red (left).
Includes bias and reconstruction error as separate bars.
"""
labels = []
values = []
colors = []
# Add positive contributors (largest first)
for feat in top_positive:
idx = feat["feature_idx"]
desc = ""
if descriptions and idx in descriptions:
desc = descriptions[idx][:40]
labels.append(f"#{idx}: {desc}")
values.append(feat["contribution"])
colors.append("#2196F3")
# Add negative contributors (most negative first)
for feat in top_negative:
idx = feat["feature_idx"]
desc = ""
if descriptions and idx in descriptions:
desc = descriptions[idx][:40]
labels.append(f"#{idx}: {desc}")
values.append(feat["contribution"])
colors.append("#F44336")
# Add bias and error
labels.append("SAE bias")
values.append(bias)
colors.append("#9E9E9E")
labels.append("Reconstruction error")
values.append(error)
colors.append("#757575")
fig = go.Figure(
data=go.Bar(
y=labels,
x=values,
orientation="h",
marker_color=colors,
hovertemplate="<b>%{y}</b><br>Contribution: %{x:.4f}<extra></extra>",
)
)
fig.update_layout(
title=f'Feature contributions to "{target_token}" (total logit: {total_logit:.2f})',
xaxis_title="Logit Contribution",
height=max(400, 30 * len(labels) + 100),
margin=dict(l=250, r=20, t=60, b=40),
yaxis=dict(autorange="reversed"),
)
return fig
def create_logit_decomposition_summary(
sae_explained: float,
bias: float,
error: float,
total: float,
) -> go.Figure:
"""Create a stacked bar chart showing SAE-explained vs bias vs error portions."""
feature_sum = sae_explained - bias # isolate pure feature contributions
labels = ["Feature contributions", "SAE bias", "Reconstruction error"]
values = [feature_sum, bias, error]
bar_colors = ["#2196F3", "#9E9E9E", "#757575"]
fig = go.Figure(
data=go.Bar(
x=labels,
y=values,
marker_color=bar_colors,
text=[f"{v:.3f}" for v in values],
textposition="auto",
)
)
gap = total - (feature_sum + bias + error)
fig.update_layout(
title=f"Logit Decomposition (total: {total:.3f}, gap: {gap:.4f})",
yaxis_title="Logit Value",
height=350,
margin=dict(l=60, r=20, t=60, b=40),
)
return fig
|