| """Chart: Beam's reasoning-effort dial next to Dory-Beam-4B's, on the evaluation set. |
| |
| .venv/bin/python -I scripts/make_effort_chart.py # -> release/model/assets/beam_effort_dial.png (+ dataset copy) |
| |
| Accuracy vs mean output tokens per answer, for three groups of benchmarks: all six, AIME 2024 + 2025 (hard), and |
| MATH-500 + AMC 2023 (routine). Beam (teacher) in categorical slot 2 (orange, as on the Space), Dory-Beam-4B in slot 1 |
| (blue), base Dory as a de-emphasised gray point. Reads data/beam_effort/summary.json from scripts/beam_effort_study.py. |
| """ |
|
|
| import json |
| import shutil |
| import sys |
| from pathlib import Path |
|
|
| import matplotlib |
|
|
| matplotlib.use("Agg") |
| import matplotlib.pyplot as plt |
| from matplotlib.ticker import FuncFormatter |
|
|
| ROOT = Path(__file__).resolve().parent.parent |
| sys.path.insert(0, str(ROOT / "scripts")) |
| from make_chart import SURFACE, INK, INK2, MUTED, GRID, AXIS, BLUE, ORANGE, BASE |
|
|
| OUT = ROOT / "release" / "model" / "assets" / "beam_effort_dial.png" |
| GROUPS = [("All six benchmarks", "all"), ("AIME 2024 + 2025", "aime")] |
| |
| OFFSETS = {"minimal": ((8, -3), "left"), "low": ((8, -3), "left"), "medium": ((8, -3), "left"), |
| "high": ((-8, 4), "right"), "xhigh": ((8, 4), "left"), "max": ((8, -11), "left")} |
| DORY_OFFSETS = {"low": ((-8, -3), "right"), "medium": ((8, -3), "left"), "high": ((8, -3), "left")} |
| BEAM_LEVELS = ["minimal", "low", "medium", "high", "xhigh", "max"] |
| SUMMARY = ROOT / "data" / "beam_effort" / "summary.json" |
|
|
|
|
| def main(): |
| chart = json.load(open(SUMMARY))["chart"] |
| k = lambda p: (p[0] / 1000, p[1]) |
| plt.rcParams.update({"font.family": ["Helvetica Neue", "Helvetica", "Arial", "DejaVu Sans"], |
| "font.size": 10, "axes.edgecolor": AXIS, "axes.labelcolor": INK2, |
| "xtick.color": MUTED, "ytick.color": MUTED, "text.color": INK}) |
| fig, axes = plt.subplots(1, 2, figsize=(11, 5.6), facecolor=SURFACE) |
| fig.subplots_adjust(left=0.07, right=0.98, top=0.71, bottom=0.12, wspace=0.18) |
| for ax, (title, g) in zip(axes, GROUPS): |
| ax.set_facecolor(SURFACE) |
| for side in ("top", "right"): |
| ax.spines[side].set_visible(False) |
| ax.grid(True, color=GRID, linewidth=1, zorder=0) |
| ax.tick_params(length=0, pad=4) |
| series = [ |
| ("Beam-1.0-501B (teacher)", ORANGE, [(lv, k(chart[g]["beam"][lv])) for lv in BEAM_LEVELS]), |
| ("Dory-Beam-4B (student) · 2 loops", BLUE, |
| [(lv, k(chart[g]["dory_beam"][lv])) for lv in ("low", "medium", "high")]), |
| ("Dory1.0-4B (base) · 2 loops", BASE, [("base", k(chart[g]["base"]))]), |
| ] |
| pts = [] |
| for label, color, ps in series: |
| xs = [p[1][0] for p in ps] |
| ys = [p[1][1] for p in ps] |
| if len(ps) > 1: |
| ax.plot(xs, ys, color=color, linewidth=2, solid_capstyle="round", solid_joinstyle="round", zorder=3) |
| ax.plot(xs, ys, "o", color=color, markersize=8, markeredgecolor=SURFACE, markeredgewidth=2, zorder=4, |
| label=label) |
| for name, (x, y) in ps: |
| off, ha = (OFFSETS if color == ORANGE else DORY_OFFSETS).get(name, ((8, -3), "left")) |
| ax.annotate(name, (x, y), xytext=off, textcoords="offset points", fontsize=8.5, color=MUTED, ha=ha) |
| pts.append((x, y)) |
| ys = [p[1] for p in pts] |
| lo, hi = min(ys), max(ys) |
| pad = max(2.0, (hi - lo) * 0.12) |
| ax.set_ylim(max(0, lo - pad), min(101, hi + pad)) |
| ax.set_xlim(0, max(p[0] for p in pts) * 1.14) |
| ax.xaxis.set_major_formatter(FuncFormatter(lambda v, _: "0" if v == 0 else f"{v:g}k")) |
| ax.yaxis.set_major_formatter(FuncFormatter(lambda v, _: f"{v:g}%")) |
| ax.set_title(title, loc="left", fontsize=11.5, fontweight="semibold", color=INK, pad=8) |
| ax.set_xlabel("mean output tokens per answer", fontsize=9.5, color=INK2) |
| axes[0].set_ylabel("accuracy", fontsize=9.5, color=INK2) |
| fig.text(0.07, 0.95, "Beam's reasoning-effort dial, and the one Dory-Beam-4B learned from it", |
| fontsize=16, fontweight="semibold", color=INK) |
| fig.text(0.07, 0.915, "Same 1,836 evaluation answers at every level; labels give the reasoning effort. Beam: self-hosted " |
| "NVFP4,\ntemperature 0.7, up to 28k tokens. Dory: temperature 1.0, up to 64k.", |
| fontsize=10, color=INK2, va="top") |
| handles, labels = axes[0].get_legend_handles_labels() |
| fig.legend(handles, labels, loc="upper left", bbox_to_anchor=(0.063, 0.83), ncol=3, frameon=False, |
| fontsize=10, handletextpad=0.4, columnspacing=1.6) |
| OUT.parent.mkdir(parents=True, exist_ok=True) |
| fig.savefig(OUT, dpi=200, facecolor=SURFACE) |
| dst = ROOT / "release" / "dataset" / "assets" |
| dst.mkdir(parents=True, exist_ok=True) |
| shutil.copy(OUT, dst / OUT.name) |
| print("wrote", OUT) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
