Spaces:
Sleeping
Sleeping
File size: 35,010 Bytes
cb8a7e5 5d228c2 cb8a7e5 5d228c2 0f312db cb8a7e5 5d228c2 cb8a7e5 5d228c2 8f79718 5d228c2 cb8a7e5 6af579c cb8a7e5 9b97db2 cb8a7e5 406b7cb 9b97db2 cb8a7e5 6af579c cb8a7e5 406b7cb cb8a7e5 6af579c cb8a7e5 406b7cb cb8a7e5 e2019d7 cb8a7e5 e2019d7 6de59a4 cb8a7e5 6de59a4 cb8a7e5 e2019d7 cb8a7e5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 | ο»Ώ"""Page 0 - Graph Generation: Generate Attribution Graphs on Neuronpedia"""
import sys
from pathlib import Path
# Add parent directory to path
parent_dir = Path(__file__).parent.parent.parent
if str(parent_dir) not in sys.path:
sys.path.insert(0, str(parent_dir))
import streamlit as st
import json
import os
from datetime import datetime
# Try to import PipelineState (optional - won't break if missing)
try:
from eda.utils.pipeline_state import PipelineState
PIPELINE_STATE_AVAILABLE = True
except ImportError:
PIPELINE_STATE_AVAILABLE = False
# Import graph generation functions
try:
from scripts.neuronpedia_graph_generation import (
generate_attribution_graph,
get_graph_stats,
load_api_key,
extract_static_metrics_from_json
)
except ImportError:
# Fallback if module is not directly importable
import importlib.util
script_path = parent_dir / "scripts" / "00_neuronpedia_graph_generation.py"
spec = importlib.util.spec_from_file_location("neuronpedia_graph_generation", script_path)
graph_gen = importlib.util.module_from_spec(spec)
spec.loader.exec_module(graph_gen)
generate_attribution_graph = graph_gen.generate_attribution_graph
get_graph_stats = graph_gen.get_graph_stats
load_api_key = graph_gen.load_api_key
extract_static_metrics_from_json = graph_gen.extract_static_metrics_from_json
st.set_page_config(page_title="Graph Generation", page_icon="π", layout="wide")
st.title("π Attribution Graph Generation")
st.info("""
1. **Generate a new attribution graph on Neuronpedia** to analyze how the model predicts the next token. \n
2. **Analyze the graph** to understand the contribution of each feature.\n
3. **Filter Features by Cumulative Influence Coverage** for downstream analysis.
""")
# ===== SIDEBAR: CONFIGURATION =====
st.sidebar.header("Configuration")
# Neuronpedia API Key
st.sidebar.subheader("Neuronpedia API")
# Try to load from environment/secrets
api_key = load_api_key()
if not api_key:
st.sidebar.warning("β οΈ Neuronpedia API Key not found")
st.sidebar.info("""
Add `NEURONPEDIA_API_KEY=your-key` to HF Secrets
or enter it below.
""")
# Allow manual input
api_key = st.sidebar.text_input(
"Enter API Key:",
type="password",
key="neuronpedia_key_input",
help="Enter your Neuronpedia API key"
)
if not api_key:
st.error("""
**Neuronpedia API Key Required!**
1. Obtain an API key from [Neuronpedia](https://www.neuronpedia.org/)
2. Enter it in the sidebar, OR
3. Add to HF Spaces Secrets (Settings β Repository secrets):
```
NEURONPEDIA_API_KEY = your-key-here
```
""")
st.stop()
else:
st.sidebar.success(f"β
API Key entered ({len(api_key)} characters)")
else:
st.sidebar.success(f"β
API Key loaded ({len(api_key)} characters)")
# Save to session_state for reuse in other pages
if api_key:
st.session_state['neuronpedia_api_key'] = api_key
# ===== SECTION: GENERATE NEW GRAPH =====
st.header("π Generate New Attribution Graph")
# INPUT PROMPT
st.subheader("1οΈβ£ Prompt Configuration")
prompt = st.text_area(
"Prompt to analyze",
value="The capital of state containing Dallas is",
height=100,
help="Enter the prompt to analyze. The model will try to predict the next token."
)
# GRAPH PARAMETERS
st.subheader("Graph Parameters")
with st.expander("Advanced configuration", expanded=False):
col1, col2 = st.columns(2)
with col1:
st.write("**Model & Source Set**")
model_id = st.selectbox(
"Model ID",
["gemma-2-2b", "gpt2-small", "gemma-2-9b"],
help="Model to analyze"
)
source_set_name = st.text_input(
"Source Set Name",
value="clt-hp", #"gemmascope-transcoder-16k",
help="Name of the SAE source set to use"
)
max_feature_nodes = st.number_input(
"Max Feature Nodes",
min_value=100,
max_value=10000,
value=5000,
step=100,
help="Maximum number of feature nodes to include"
)
with col2:
st.write("**Thresholds**")
node_threshold = st.slider(
"Node Threshold",
min_value=0.0,
max_value=1.0,
value=0.8,
step=0.05,
help="Minimum importance threshold to include a node"
)
edge_threshold = st.slider(
"Edge Threshold",
min_value=0.0,
max_value=1.0,
value=0.85,
step=0.05,
help="Minimum importance threshold to include an edge"
)
max_n_logits = st.number_input(
"Max N Logits",
min_value=1,
max_value=50,
value=10,
step=1,
help="Maximum number of logits to consider"
)
desired_logit_prob = st.slider(
"Desired Logit Probability",
min_value=0.5,
max_value=0.99,
value=0.95,
step=0.01,
help="Desired cumulative probability for logits"
)
slug = st.text_input(
"Custom slug (optional)",
value="",
help="If empty, will be generated automatically"
)
# GENERATION
st.subheader("Generation")
col1, col2 = st.columns([1, 2])
with col1:
generate_button = st.button("π Generate Graph", type="primary", use_container_width=True)
with col2:
save_locally = st.checkbox("Save locally", value=True)
# State
if 'generation_result' not in st.session_state:
st.session_state.generation_result = None
if 'static_metrics_df' not in st.session_state:
st.session_state.static_metrics_df = None
if 'extracted_graph_data' not in st.session_state:
st.session_state.extracted_graph_data = None
if 'extracted_csv_df' not in st.session_state:
st.session_state.extracted_csv_df = None
if generate_button:
if not prompt.strip():
st.error("Enter a valid prompt!")
st.stop()
progress_bar = st.progress(0)
status_text = st.empty()
try:
status_text.text("Preparing...")
progress_bar.progress(10)
status_text.text("Sending request to Neuronpedia...")
progress_bar.progress(30)
result = generate_attribution_graph(
prompt=prompt,
api_key=api_key,
model_id=model_id,
source_set_name=source_set_name,
slug=slug if slug.strip() else None,
max_n_logits=max_n_logits,
desired_logit_prob=desired_logit_prob,
node_threshold=node_threshold,
edge_threshold=edge_threshold,
max_feature_nodes=max_feature_nodes,
save_locally=save_locally,
verbose=False
)
progress_bar.progress(100)
status_text.empty()
progress_bar.empty()
# Add generation parameters to result for later use
if result['success']:
result['source_set_name'] = source_set_name
result['node_threshold'] = node_threshold
result['desired_logit_prob'] = desired_logit_prob
# Rename file to new format if saved locally (BEFORE saving to session_state)
if result['success'] and result.get('local_path') and save_locally and PIPELINE_STATE_AVAILABLE:
old_path = Path(result['local_path'])
if old_path.exists():
# Generate new filename with st1_ prefix
new_filename = PipelineState.generate_filename(
step=1,
file_type='graph',
prompt=prompt
)
new_path = old_path.parent / new_filename
# Rename file
old_path.rename(new_path)
# Update result with absolute path (AFTER rename)
result['local_path'] = str(new_path.resolve())
result['renamed_to_new_format'] = True
# Save result to session_state (with updated path)
st.session_state.generation_result = result
# Save Graph JSON to pipeline session_state for auto-loading in next steps
if result['success'] and result.get('local_path'):
try:
with open(result['local_path'], 'r', encoding='utf-8') as f:
graph_data = json.load(f)
st.session_state['pipeline_graph_json'] = {
'data': graph_data,
'filename': Path(result['local_path']).name,
'timestamp': datetime.now().isoformat()
}
except Exception as e:
# Don't break the flow if saving to pipeline state fails
pass
# Build Neuronpedia URL
if result['success']:
neuronpedia_url = (
f"https://www.neuronpedia.org/{result.get('model_id', 'gemma-2-2b')}/graph"
f"?sourceSet={result.get('source_set_name', 'clt-hp')}"
f"&slug={result.get('slug', '')}"
f"&pruningThreshold={result.get('node_threshold', 0.8)}"
f"&densityThreshold={result.get('desired_logit_prob', 0.95)}"
)
# Get the filename for display
if result.get('local_path'):
filename = Path(result['local_path']).name
st.success(f"β
Graph generated successfully: `{filename}`\n\n" f"[**Open Graph on Neuronpedia**]({neuronpedia_url})")
# Auto-download the generated graph JSON
try:
import streamlit.components.v1 as components
import base64
with open(result['local_path'], 'r', encoding='utf-8') as f:
graph_json_content = f.read()
# Encode to base64 for JavaScript
b64 = base64.b64encode(graph_json_content.encode()).decode()
# Auto-download with JavaScript
html = f"""
<script>
function downloadFile() {{
const link = document.createElement('a');
link.href = 'data:application/json;base64,{b64}';
link.download = '{filename}';
document.body.appendChild(link);
link.click();
document.body.removeChild(link);
}}
// Trigger download after a short delay
setTimeout(downloadFile, 100);
</script>
"""
components.html(html, height=50)
except Exception as e:
st.warning(f"β οΈ Could not prepare auto-download: {e}")
else:
# Generation failed
error_msg = result.get('error', 'Unknown error')
st.error(f"β Graph generation failed!\n\nError: {error_msg}")
# Show details if available
if result.get('details'):
with st.expander("Error details"):
st.code(result['details'])
except Exception as e:
progress_bar.empty()
status_text.empty()
st.error(f"Unexpected error: {str(e)}")
with st.expander("Details"):
import traceback
st.code(traceback.format_exc())
st.markdown("---")
# ===== SECTION: ANALYZE GRAPH =====
st.subheader("2οΈβ£ Analyze Graph")
# Check if we just generated a graph
just_generated = st.session_state.get('generation_result') and st.session_state.generation_result.get('success')
generated_path = st.session_state.generation_result.get('local_path') if just_generated else None
if just_generated and generated_path:
# Auto-select the just-generated graph
from pathlib import Path as PathLib
# Use the absolute path directly - we know it exists
gen_path = PathLib(generated_path)
# Store the absolute path for use later
selected_json = str(gen_path)
st.info(f"π **Ready to analyze**: `{gen_path.name}` (just generated)")
# Option to select a different file
with st.expander("π Select a different graph file", expanded=False):
json_dir = parent_dir / "output" / "graph_data"
# Create directory if it doesn't exist
if not json_dir.exists():
try:
json_dir.mkdir(parents=True, exist_ok=True)
except Exception:
pass # Silently fail in expander
if json_dir.exists():
json_files = sorted(json_dir.glob("*.json"), key=lambda x: x.stat().st_mtime, reverse=True)
if json_files:
json_options = [str(f.relative_to(parent_dir)) for f in json_files]
# Find index of generated file
default_idx = 0
try:
default_idx = json_options.index(selected_json)
except ValueError:
pass
selected_json_alt = st.selectbox(
"Select JSON file",
options=json_options,
index=default_idx,
key="alt_json_select",
help="JSON files sorted by date (most recent first)"
)
if st.button("Use this file instead"):
selected_json = selected_json_alt
st.rerun()
else:
st.info("No other graph files found in `output/graph_data/`")
else:
st.warning("Directory `output/graph_data/` not accessible")
else:
# Normal file selection (no graph just generated)
st.write("""
Extract static metrics (`node_influence`, `cumulative_influence`, `frac_external_raw`) from an existing graph.
""")
json_dir = parent_dir / "output" / "graph_data"
# Create directory if it doesn't exist
if not json_dir.exists():
try:
json_dir.mkdir(parents=True, exist_ok=True)
except Exception as e:
st.warning(f"β οΈ Could not create directory: {e}")
if json_dir.exists():
json_files = sorted(json_dir.glob("*.json"), key=lambda x: x.stat().st_mtime, reverse=True)
if json_files:
# Use relative paths for display
json_options = [str(f.relative_to(parent_dir)) for f in json_files]
selected_json = st.selectbox(
"Select JSON file",
options=json_options,
help="JSON files sorted by date (most recent first)"
)
else:
st.warning("No JSON files found in `output/graph_data/`")
selected_json = None
else:
st.warning("Directory `output/graph_data/` not found")
selected_json = None
# Show file info and analysis button if we have a selected file
if selected_json:
# Handle both absolute and relative paths
file_path = Path(selected_json)
if not file_path.is_absolute():
file_path = parent_dir / selected_json
# Check if file exists - if not, try to use session_state data
file_exists = file_path.exists()
# If file doesn't exist but we have data in session_state (just generated), use that
use_session_data = False
if not file_exists and just_generated and 'pipeline_graph_json' in st.session_state:
graph_data = st.session_state['pipeline_graph_json']['data']
use_session_data = True
st.info("π¦ Using graph data from current session (file system is temporary on HF Spaces)")
elif not file_exists:
st.error(f"β File not found: `{file_path.name}`")
st.warning("The file may have been moved or renamed. Please refresh the page or select another file.")
st.stop()
# Get file stats and metadata
if use_session_data:
# Use data from session_state
file_size = len(json.dumps(graph_data)) / 1024 / 1024 # Approximate size
file_time = datetime.fromisoformat(st.session_state['pipeline_graph_json']['timestamp'])
num_nodes = len(graph_data.get('nodes', []))
num_links = len(graph_data.get('links', []))
model_id = graph_data.get('metadata', {}).get('model_id', 'N/A')
else:
# Use file on disk
file_size = file_path.stat().st_size / 1024 / 1024
file_time = datetime.fromtimestamp(file_path.stat().st_mtime)
# Load JSON to extract graph metadata
try:
with open(file_path, 'r', encoding='utf-8') as f:
graph_metadata = json.load(f)
num_nodes = len(graph_metadata.get('nodes', []))
num_links = len(graph_metadata.get('links', []))
model_id = graph_metadata.get('metadata', {}).get('model_id', 'N/A')
except Exception:
num_nodes = None
num_links = None
model_id = None
# Display file info and graph metadata
col1, col2, col3 = st.columns(3)
with col1:
st.metric("Size", f"{file_size:.2f} MB")
with col2:
st.metric("Date", file_time.strftime("%Y-%m-%d %H:%M"))
with col3:
st.metric("Name", file_path.name[:20] + "...")
if num_nodes is not None and num_links is not None and model_id is not None:
col4, col5, col6 = st.columns(3)
with col4:
st.metric("Nodes", num_nodes)
with col5:
st.metric("Links", num_links)
with col6:
st.metric("Model", model_id)
# Extract button
button_label = "π Analyze This Graph" if just_generated else "π Analyze Graph"
if st.button(button_label, key="extract_existing", type="primary"):
try:
with st.spinner("Extracting metrics..."):
# Use graph_data from session_state if available, otherwise load from file
if use_session_data:
# Already have graph_data from session_state
pass
else:
# Load from file
json_full_path = str(parent_dir / selected_json)
with open(json_full_path, 'r', encoding='utf-8') as f:
graph_data = json.load(f)
csv_output_path = str(parent_dir / "output" / "graph_feature_static_metrics.csv")
df = extract_static_metrics_from_json(
graph_data,
output_path=csv_output_path,
verbose=False
)
# Save in session_state to persist across reruns
st.session_state.extracted_graph_data = graph_data
st.session_state.extracted_csv_df = df
st.session_state.analysis_performed = True
st.success(f"β
CSV generated: `{csv_output_path}`")
st.info("π Scroll down to see interactive visualizations")
except Exception as e:
st.error(f"β Error: {str(e)}")
st.markdown("---")
# ===== EXTRACTED DATA VISUALIZATION (persists across reruns) =====
if st.session_state.extracted_graph_data is not None and st.session_state.extracted_csv_df is not None:
graph_data = st.session_state.extracted_graph_data
df = st.session_state.extracted_csv_df
# Only show if analysis was performed
if st.session_state.get('analysis_performed', False):
st.header("Extracted Data Analysis")
# CSV Metrics
col1, col2, col3, col4, col5 = st.columns(5)
with col1:
st.metric("Features", len(df))
with col2:
st.metric("Unique Tokens", df['ctx_idx'].nunique())
with col3:
st.metric("Mean Activation", f"{df['activation'].mean():.3f}")
with col4:
# Use node_influence (marginal influence) for total sum
st.metric("Sum Node Infl", f"{df['node_influence'].sum():.2f}")
with col5:
st.metric("Mean Frac Ext", f"{df['frac_external_raw'].mean():.3f}")
with st.expander("View Complete Dataframe", expanded=False):
st.dataframe(df, use_container_width=True, height=600)
# Scatter plot: Layer vs Context Position with Influence
# Prepare data from JSON for scatter plot
if 'nodes' in graph_data:
import pandas as pd
import plotly.express as px
# Extract prompt_tokens from metadata to map ctx_idx -> token
prompt_tokens = graph_data.get('metadata', {}).get('prompt_tokens', [])
# Scatter plot visualization with filter
from eda.utils.graph_visualization import create_scatter_plot_with_filter
filtered_features = create_scatter_plot_with_filter(graph_data)
# Save filtered_features for export section
if filtered_features is not None and len(filtered_features) > 0:
st.session_state.filtered_features_export = filtered_features
# ===== SUMMARY CHARTS: COVERAGE AND STRENGTH =====
# Only show if analysis was performed
if st.session_state.get('analysis_performed', False):
# Data source: prefer extracted data, otherwise last generated graph
graph_data_for_plots = None
if st.session_state.get('extracted_graph_data') is not None:
graph_data_for_plots = st.session_state.extracted_graph_data
elif st.session_state.get('generation_result') is not None and st.session_state.generation_result.get('success'):
graph_data_for_plots = st.session_state.generation_result.get('graph_data')
if graph_data_for_plots is not None and 'nodes' in graph_data_for_plots:
with st.expander("Summary Charts: Coverage and Strength", expanded=False):
import pandas as pd
import plotly.express as px
import numpy as np
nodes_df = pd.DataFrame(graph_data_for_plots['nodes'])
is_feature = nodes_df['node_id'].astype(str).str[0].str.isdigit() & nodes_df['node_id'].astype(str).str.contains('_')
feat_nodes = nodes_df.loc[is_feature].copy()
if len(feat_nodes) == 0:
st.warning("No features found in current data.")
else:
# Add slider to filter (reuse same logic as create_scatter_plot_with_filter)
max_influence = feat_nodes['influence'].max()
st.markdown("### Filter Features by Cumulative Influence")
st.info(f"""
**Use the slider to filter the charts below** based on cumulative influence coverage (0-{max_influence:.2f}).
Summary charts will show only features with `influence <= threshold`.
""")
# Check if main slider already exists (from create_scatter_plot_with_filter)
# If it exists, use it, otherwise create a new one
slider_key = "cumulative_slider_summary"
if "cumulative_slider_main" in st.session_state:
# Reuse main slider value
cumulative_threshold_summary = st.session_state.cumulative_slider_main
st.info(f"Synchronized with main slider: threshold = {cumulative_threshold_summary:.4f}")
else:
# Create separate slider
cumulative_threshold_summary = st.slider(
"Cumulative Influence Threshold (summary charts)",
min_value=0.0,
max_value=float(max_influence),
value=float(max_influence),
step=0.01,
key=slider_key,
help=f"Keep only features with influence <= threshold. Range: 0.0 - {max_influence:.2f}"
)
# Apply filter
feat_nodes_filtered = feat_nodes[feat_nodes['influence'] <= cumulative_threshold_summary].copy()
if len(feat_nodes_filtered) == 0:
st.warning("No features match the current filter. Increase the threshold.")
else:
# Show filter statistics
col1, col2, col3 = st.columns(3)
with col1:
st.metric("Total Features", len(feat_nodes))
with col2:
st.metric("Filtered Features", len(feat_nodes_filtered))
with col3:
pct = (len(feat_nodes_filtered) / len(feat_nodes) * 100) if len(feat_nodes) > 0 else 0
st.metric("% Kept", f"{pct:.1f}%")
st.markdown("---")
# Calculate n_ctx and statistics per feature
feat_nodes_filtered['feature_key'] = feat_nodes_filtered['node_id'].str.rsplit('_', n=1).str[0]
cov = (
feat_nodes_filtered.groupby('feature_key')['ctx_idx'].nunique()
.rename('n_ctx').reset_index()
)
per_feat = (
feat_nodes_filtered.groupby('feature_key')
.agg(mean_influence=('influence','mean'),
mean_activation=('activation','mean'))
.reset_index()
)
per_feat_cov = per_feat.merge(cov, on='feature_key', how='left')
nodes_with_cov = feat_nodes_filtered.merge(cov, on='feature_key', how='left')
# Chart 1: Coverage (Histogram + ECDF)
st.subheader("Feature Coverage (n_ctx)")
c1, c2 = st.columns(2)
with c1:
fig_hist = px.histogram(cov, x='n_ctx', color_discrete_sequence=['#4C78A8'])
fig_hist.update_layout(title='n_ctx distribution per feature',
xaxis_title='Number of unique ctx_idx',
yaxis_title='Number of features')
st.plotly_chart(fig_hist, use_container_width=True)
with c2:
fig_ecdf = px.ecdf(cov, x='n_ctx', color_discrete_sequence=['#F58518'])
fig_ecdf.update_layout(title='n_ctx ECDF',
xaxis_title='Number of unique ctx_idx',
yaxis_title='Cumulative fraction')
st.plotly_chart(fig_ecdf, use_container_width=True)
# Chart 2: Strength vs Coverage (Activation vs n_ctx and Scatter mean)
st.subheader("Strength vs Coverage")
c3, c4 = st.columns(2)
with c3:
fig_violin = px.violin(nodes_with_cov, x='n_ctx', y='activation', box=True, points=False)
fig_violin.update_layout(title='Activation per n_ctx',
xaxis_title='n_ctx (feature)',
yaxis_title='Activation (node)')
st.plotly_chart(fig_violin, use_container_width=True)
with c4:
fig_scatter = px.scatter(per_feat_cov, x='mean_activation', y='mean_influence',
color='n_ctx', size='n_ctx', hover_data=['feature_key'],
color_continuous_scale='Viridis')
# Correlations for subtitle
if len(per_feat_cov) >= 2:
pearson = float(per_feat_cov['mean_activation'].corr(per_feat_cov['mean_influence'], method='pearson'))
spearman = float(per_feat_cov['mean_activation'].corr(per_feat_cov['mean_influence'], method='spearman'))
fig_scatter.update_layout(title=f'Mean activation vs mean influence<br>(r={pearson:.2f}, rho={spearman:.2f})')
else:
fig_scatter.update_layout(title='Mean activation vs mean influence')
fig_scatter.update_layout(xaxis_title='Mean activation (per feature)',
yaxis_title='Mean influence (per feature)')
st.plotly_chart(fig_scatter, use_container_width=True)
# Quick insights
with st.expander("Insights from charts", expanded=False):
# Calculate key statistics
top_n_ctx = cov['n_ctx'].max()
n_top = len(cov[cov['n_ctx'] == top_n_ctx])
top_features = cov[cov['n_ctx'] == top_n_ctx]['feature_key'].tolist()
st.markdown(f"""
**Coverage (n_ctx)**:
- {len(cov)} unique features in filtered dataset
- {n_top} features present in all {top_n_ctx} contexts
- Multi-context features ({top_n_ctx}): {', '.join([f'`{f}`' for f in top_features[:5]])}
**Strength vs Coverage**:
- Activation-influence correlation: **r={pearson:.2f}** (Pearson), **rho={spearman:.2f}** (Spearman)
- {"Negative correlation: features with high activation tend to have low influence" if pearson < -0.2 else "Weak or positive correlation between activation and influence"}
""")
# Group statistics
if len(nodes_with_cov) > 0:
g1 = nodes_with_cov[nodes_with_cov['n_ctx'] == 1]
g_multi = nodes_with_cov[nodes_with_cov['n_ctx'] >= 5]
if len(g1) > 0 and len(g_multi) > 0:
st.markdown(f"""
**Group comparison**:
- n_ctx=1: {len(g1)} nodes, mean_activation={g1['activation'].mean():.2f}, mean_influence={g1['influence'].mean():.3f}
- n_ctx>=5: {len(g_multi)} nodes, mean_activation={g_multi['activation'].mean():.2f}, mean_influence={g_multi['influence'].mean():.3f}
""")
# ===== EXPORT SELECTED FEATURES =====
if st.session_state.get('analysis_performed', False) and st.session_state.get('filtered_features_export') is not None:
filtered_features = st.session_state.filtered_features_export
if len(filtered_features) > 0:
st.markdown("---")
st.subheader("Export Selected Features")
# Convert dataframe to format [{"layer": X, "index": Y}, ...]
# Remove duplicates using set of tuples (layer, feature)
unique_features = {
(int(row['layer']), int(row['feature']))
for _, row in filtered_features.iterrows()
}
# Convert to sorted list of dicts
features_export = [
{"layer": layer, "index": feature}
for layer, feature in sorted(unique_features)
]
# Also extract selected node_ids (for subgraph upload)
node_ids_export = sorted(filtered_features['id'].unique().tolist())
# Create complete export with features AND node_ids
export_data = {
"features": features_export,
"node_ids": node_ids_export,
"metadata": {
"n_features": len(features_export),
"n_nodes": len(node_ids_export),
"cumulative_threshold": st.session_state.get('cumulative_slider_main', None),
"exported_at": datetime.now().isoformat()
}
}
# Statistics
col1, col2, col3 = st.columns(3)
with col1:
st.metric("Unique Features", len(features_export))
with col2:
st.metric("Selected Nodes", len(node_ids_export))
with col3:
st.metric("Unique Layers", len({f['layer'] for f in features_export}))
# Save to pipeline session_state for auto-loading in next steps
st.session_state['pipeline_selected_nodes'] = {
'data': export_data,
'filename': f"st1_feat_node_subset_{datetime.now().strftime('%Y%m%d_%H%M%S')}.json",
'timestamp': datetime.now().isoformat()
}
# Download JSON (complete format)
st.download_button(
label="π₯ Download Features+Nodes Subset",
data=json.dumps(export_data, indent=2, ensure_ascii=False),
file_name="selected_features_with_nodes.json",
mime="application/json",
help="Complete format with features and node_ids (for Node Grouping + Probe Prompts + batch_get_activations.py)",
use_container_width=True,
type="primary"
)
# LEGACY BUTTON (hidden - all tools now support complete format)
# with col_legacy:
# st.download_button(
# label="Download Features JSON (legacy)",
# data=json.dumps(features_export, indent=2, ensure_ascii=False),
# file_name="selected_features.json",
# mime="application/json",
# help="Legacy format (features only, compatible with batch_get_activations.py)"
# )
# Preview
with st.expander("Preview Complete Export", expanded=False):
st.json({
"features": features_export[:5],
"node_ids": node_ids_export[:10],
"metadata": export_data["metadata"]
})
# ===== FOOTER =====
st.sidebar.markdown("---")
st.sidebar.subheader("Info")
st.sidebar.markdown("""
**Attribution Graph**: visualizes how SAE features contribute to predictions.
**Elements**:
- Embedding nodes: input tokens
- Feature nodes: SAE latents
- Logit nodes: predicted tokens
""")
st.sidebar.caption("Powered by Neuronpedia API")
|