from __future__ import annotations from pathlib import Path import gradio as gr import pandas as pd import plotly.graph_objects as go DATA = pd.read_parquet( Path(__file__).resolve().parent / "data" / "feature_exemplars.parquet" ) FEATURES = sorted(DATA["feature"].unique().tolist()) def inspect_feature(feature: int) -> tuple[go.Figure, dict]: feature = int(feature) rows = DATA[DATA["feature"] == feature].sort_values("rank") figure = go.Figure( go.Bar( x=rows["token"].tolist(), y=rows["activation"].tolist(), marker_color="#a78bfa", customdata=rows["token_id"].tolist(), hovertemplate="token=%{x}
id=%{customdata}
activation=%{y:.3f}", ) ) figure.update_layout( title=f"Feature {feature}: top held-out token activations", xaxis_title="BPE token", yaxis_title="Sparse feature activation", template="plotly_dark", ) return figure, { "feature": feature, "heldout_firing_rate": round(float(rows["firing_rate"].iloc[0]), 6), "top_tokens": rows["token"].tolist(), "warning": "Token exemplars are clues, not causal concept labels.", } with gr.Blocks(title="SNIP Scope") as demo: gr.Markdown( "# SNIP Scope\n" "Explore top held-out token activations for 384 sparse features learned " "from the SNIP transformer's final hidden layer." ) feature = gr.Slider(0, max(FEATURES), 0, step=1, label="Sparse feature") run = gr.Button("Inspect feature", variant="primary") exemplars = gr.Plot() details = gr.JSON() run.click(inspect_feature, feature, [exemplars, details]) demo.load(inspect_feature, feature, [exemplars, details]) if __name__ == "__main__": demo.launch()