File size: 5,390 Bytes
f227f52 | 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 | #!/usr/bin/env python3
"""Render the top-three non-DAPO reward curves with base-model references."""
from pathlib import Path
import matplotlib.pyplot as plt
import pandas as pd
ARTIFACTS = Path(__file__).resolve().parent
INPUT = ARTIFACTS / "top-three-v49-no-dapo-real-reward.csv"
OUTPUT_STEM = ARTIFACTS / "top-three-v49-no-dapo-real-reward"
# Qwen3-Coder-30B-A3B-Instruct TaskTrove v4.6 evaluation. The paired pass@1
# values use the same complete 16-attempt cohorts as the pass@16 estimates.
REFERENCES = {
"Agent": {"pass@1": 0.48936170212765956, "pass@16": 0.608843537414966},
}
def display_label(job: str) -> str:
if "entropy3e4-lr2-g16" in job:
return "RLOO, entropy 0.0003, LR 2e-6, group 16"
if "rloo-entropy3e3" in job:
return "RLOO, entropy 0.003, LR 1e-6"
if "nodapo-lr2" in job:
return "RLOO, LR 2e-6"
if "nodapo-lr4" in job:
return "RLOO, LR 4e-6"
raise ValueError(f"No concise display label defined for {job}")
def main() -> None:
frame = pd.read_csv(INPUT)
run_rows = frame[frame["reward_scope"] != "reported_v46_reference"].copy()
summaries = (
run_rows.groupby(["dataset", "job"], as_index=False)
.agg(step_count=("step", "nunique"), peak=("real_reward", "max"))
.query("step_count >= 5")
.sort_values(["dataset", "peak"], ascending=[True, False])
.groupby("dataset", as_index=False)
.head(3)
)
run_rows = run_rows.merge(summaries[["dataset", "job"]], on=["dataset", "job"], how="inner")
run_rows["label"] = run_rows["job"].map(display_label)
reference_rows = []
for dataset, values in REFERENCES.items():
for metric, value in values.items():
reference_rows.append(
{
"dataset": dataset,
"job": "Qwen3-Coder-30B-A3B-Instruct",
"label": f"Reported v4.6 {metric}",
"step": "",
"real_reward": value,
"reward_scope": "reported_v46_reference",
}
)
pd.concat([run_rows, pd.DataFrame(reference_rows)], ignore_index=True).to_csv(INPUT, index=False)
plt.rcParams.update(
{
"font.family": "DejaVu Sans",
"font.size": 12,
"axes.titlesize": 18,
"axes.titleweight": "bold",
"axes.labelsize": 13,
"legend.fontsize": 10.5,
}
)
fig, axes = plt.subplots(1, 2, figsize=(15.5, 11.5), sharey=True)
colors = ["#0072B2", "#D55E00", "#009E73", "#CC79A7", "#E69F00"]
for ax, dataset in zip(axes, ("Instruction-following", "Agent"), strict=True):
panel = run_rows[run_rows["dataset"] == dataset]
order = (
panel.groupby(["job", "label"], as_index=False)["real_reward"]
.max()
.sort_values("real_reward", ascending=False)
)
for color, row in zip(colors, order.itertuples(index=False), strict=False):
series = panel[panel["job"] == row.job].sort_values("step")
ax.plot(
series["step"],
series["real_reward"],
color=color,
marker="o",
markersize=4.5,
linewidth=2.2,
label=f"{row.label} (peak {row.real_reward:.3f})",
zorder=3,
)
peak = series.loc[series["real_reward"].idxmax()]
ax.scatter(
[peak["step"]],
[peak["real_reward"]],
marker="D",
s=58,
color=color,
edgecolor="white",
linewidth=0.8,
zorder=4,
)
if dataset in REFERENCES:
reference_styles = (("pass@1", (0, (6, 4))), ("pass@16", (0, (2, 2))))
for metric, dash in reference_styles:
value = REFERENCES[dataset][metric]
ax.axhline(
value,
color="#4D4D4D",
linestyle=dash,
linewidth=1.7,
alpha=0.9,
label=f"Base model {metric} ({value:.3f}, v4.6)",
zorder=2,
)
ax.set_title(dataset, pad=12)
ax.set_xlabel("Training step")
ax.set_xlim(0.15, max(18.85, float(panel["step"].max()) + 0.85))
ax.set_ylim(0, 0.82)
ax.grid(True, color="#CFD5DC", linewidth=0.7, alpha=0.65)
ax.set_axisbelow(True)
ax.legend(
loc="upper center",
bbox_to_anchor=(0.5, -0.13),
frameon=False,
ncol=1,
)
axes[0].set_ylabel("Verifier outcome reward")
fig.suptitle("Top three v4.9 non-DAPO runs by peak real reward", fontsize=21, fontweight="bold", y=0.98)
fig.text(
0.5,
0.935,
"Unfiltered batch reward; runs have at least five completed steps. "
"Agent-panel dashed references are Qwen3-Coder base-model TaskTrove v4.6 evaluations.",
ha="center",
fontsize=13,
color="#444444",
)
fig.tight_layout(rect=(0, 0.20, 1, 0.90), w_pad=3.0)
fig.savefig(OUTPUT_STEM.with_suffix(".png"), dpi=180, bbox_inches="tight")
fig.savefig(OUTPUT_STEM.with_suffix(".svg"), bbox_inches="tight")
if __name__ == "__main__":
main()
|