"""Render both companion figures using only the verified replay JSON. Requires Matplotlib.""" import json from pathlib import Path import matplotlib matplotlib.use("Agg") import matplotlib.pyplot as plt from matplotlib.ticker import PercentFormatter ROOT = Path(__file__).resolve().parents[1] OUT = ROOT / "docs" / "figures" OUT.mkdir(exist_ok=True) models = json.loads((ROOT / "reference/replay.json").read_text(encoding="utf-8"))["models"] qwen, smol = (models[key] for key in ("Qwen2.5-3B-Instruct", "SmolLM2-1.7B-Instruct")) green, blue, ink, grid = "#15624f", "#415c82", "#172c29", "#dde5e0" plt.rcParams.update({"font.family": "DejaVu Sans", "font.size": 12, "text.color": ink, "axes.labelcolor": ink, "xtick.color": ink, "ytick.color": ink, "axes.spines.top": False, "axes.spines.right": False, "axes.spines.left": False, "axes.spines.bottom": False}) def canvas(title): fig, ax = plt.subplots(figsize=(8, 4.6), dpi=160) fig.subplots_adjust(left=.32, right=.905, top=.69, bottom=.18) fig.text(.07, .91, title, fontsize=16, fontweight="bold") fig.legend(handles=[plt.Line2D([], [], marker="o", linestyle="", color=green, label="Qwen2.5-3B"), plt.Line2D([], [], marker="s", linestyle="", color=blue, label="SmolLM2-1.7B")], loc="upper left", bbox_to_anchor=(.057, .86), frameon=False, ncol=2, fontsize=11) ax.tick_params(axis="both", length=0, pad=8) ax.xaxis.grid(color=grid, zorder=0) ax.set_axisbelow(True) return fig, ax def plot_pair(ax, y, a, b): ax.plot([min(a,b), max(a,b)], [y,y], lw=2.0, color=grid, zorder=1) ax.plot(b, y+.095, "s", ms=7, color=blue, zorder=3) ax.plot(a, y-.095, "o", ms=8, color=green, zorder=4) def export(fig, name): fig.savefig(OUT / name, dpi=160, facecolor="white", metadata={"Software":"PAIRSAFE research companion"}) plt.close(fig) fig, ax = canvas("Three ways to score the same 150 pairs") fields = [("Ordering (PSA)", "psa"), ("Best shared threshold", "oracle_pcda"), ("Held-out threshold", "calibrated_pcda")] for i, (_, key) in enumerate(fields): plot_pair(ax, i, qwen["contrastive_margin"][key], smol["contrastive_margin"][key]) ax.set(yticks=range(3), yticklabels=[s for s,_ in fields], ylim=(2.55,-.55), xlim=(.70,1.015), xticks=[.70,.75,.80,.85,.90,.95,1.0], xlabel="Share of matched pairs") ax.xaxis.set_major_formatter(PercentFormatter(1, decimals=0)) export(fig, "paired-decisions.png") fig, ax = canvas("A shared threshold within each category") fields = [("Definitions","definitions"), ("Figurative language","figurative_language"), ("Historical events","historical_events"), ("Homonyms","homonyms"), ("Safe contexts","safe_contexts"), ("Safe targets","safe_targets")] for i, (_, key) in enumerate(fields): plot_pair(ax, i, qwen["by_type"][key]["oracle_pcda"], smol["by_type"][key]["oracle_pcda"]) ax.set(yticks=range(6), yticklabels=[s for s,_ in fields], ylim=(5.55,-.55), xlim=(.66,1.02), xticks=[.68,.76,.84,.92,1.0], xlabel="Oracle-PCDA across 25 pairs per category") ax.xaxis.set_major_formatter(PercentFormatter(1, decimals=0)) export(fig, "category-thresholds.png")