docs: use one Quarto source for README and Pages

2026-06-27 16:46:08 +08:00 · 2026-06-25 13:06:12 +08:00
parent 024fb3d545
commit cfcb57b9ce
20 changed files with 533 additions and 2000 deletions
@@ -1,112 +0,0 @@
---
-title: Persona Steering Template Library
-format:
-  html:
-    toc: true
-    code-fold: true
-jupyter: python3
-execute:
-  echo: false
-  warning: false
-  message: false
---
-
-```{python}
-from pathlib import Path
-import html
-import json
-import sys
-import textwrap
-
-import plotly.graph_objects as go
-
-ROOT = Path.cwd().parent
-sys.path.insert(0, str(ROOT / "scripts"))
-```
-
-This page is the interactive companion to the README. Use hover labels to inspect
-the refusal-pole probe without forcing the README plot to carry every label.
-
-## Refusal-Pole Probe
-
-```{python}
-summary_path = ROOT / "out/model_matrix/refusal_probe_seed24_n1_template_model_summary.jsonl"
-rows = [json.loads(line) for line in summary_path.read_text().splitlines() if line.strip()]
-
-
-def wrap_tooltip_text(text: str, width: int = 56) -> str:
-    escaped = html.escape(" ".join(text.split()))
-    return "<br>".join(
-        textwrap.wrap(escaped, width=width, break_long_words=True, break_on_hyphens=False))
-
-
-plot_rows = []
-for i, row in enumerate(rows, start=1):
-    plot_rows.append({
-        "rank": i,
-        "template": row["template"],
-        "on_axis": min(1.0, max(0.0, row["axis_delta_mean"] / 8.0)),
-        "off_axis": min(1.0, max(0.0, (row["off_axis_problem_mean"] - 1.0) / 6.0)),
-        "score_p25": row["score_p25"],
-        "score_t": row["score_t"],
-        "score_mean": row["score_mean"],
-        "score_std": row["score_std"],
-        "pass": row["strict_pass_rate_mean"],
-        "echo": row["persona_echo_rate_mean"],
-        "refusal": row["refusal_or_ai_break_rate_mean"],
-    })
-
-hover = [
-    "<br>".join([
-        f"<b>{wrap_tooltip_text(row['template'])}</b>",
-        f"rank: {row['rank']}",
-        f"score t: {row['score_t']:.2f}",
-        f"score p25: {row['score_p25']:.2f}",
-        f"score mean: {row['score_mean']:.2f}",
-        f"score std: {row['score_std']:.2f}",
-        f"strict pass: {row['pass']:.3f}",
-        f"echo: {row['echo']:.3f}",
-        f"refusal: {row['refusal']:.3f}",
-        f"on-axis: {row['on_axis']:.3f}",
-        f"off-axis: {row['off_axis']:.3f}",
-    ])
-    for row in plot_rows
-]
-
-fig = go.Figure(
-    data=go.Scatter(
-        x=[row["on_axis"] for row in plot_rows],
-        y=[row["off_axis"] for row in plot_rows],
-        mode="markers",
-        text=hover,
-        hovertemplate="%{text}<extra></extra>",
-        marker={
-            "size": 9,
-            "color": [row["pass"] for row in plot_rows],
-            "colorscale": "Greys",
-            "showscale": True,
-            "colorbar": {"title": "strict pass"},
-            "line": {"width": 0},
-        },
-    )
-)
-fig.update_layout(
-    autosize=True,
-    height=680,
-    yaxis={"range": [-0.02, 1.02]},
-    xaxis={"range": [-0.02, 1.02]},
-    template="plotly_white",
-    margin={"l": 70, "r": 20, "t": 20, "b": 70},
-    xaxis_title="template on-axis movement, higher is better",
-    yaxis_title="template off-axis confounding, lower is better",
-)
-fig.show()
-```
-
-Each point is one template, averaged over two refusal-probe axes and four clean
-model artifacts. Lower-right is better: more intended-axis movement with less
-off-axis confounding.
-
-## Static SVG
-
-![Static refusal-pole probe](out/model_matrix/refusal_probe_seed24_n1_model_matrix.svg)