File size: 7,566 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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
#!/usr/bin/env python3
"""Plot the fixed-validation curves for the software-transfer arms."""

from pathlib import Path

import matplotlib.pyplot as plt
import pandas as pd
from matplotlib.lines import Line2D


ARTIFACTS = Path(__file__).resolve().parent
INPUT = ARTIFACTS / "software-transfer-fixed-validation.csv"
OUTPUT = ARTIFACTS / "software-transfer-fixed-validation"

SERIES = {
    "factorial_constdenom_g4_entropy_0.00003_lr_2e-6": ("constant_denominator", 4),
    "factorial_sequence_mean_g8_entropy_0.00003_lr_2e-6": ("sequence_mean", 8),
    "factorial_sequence_mean_g16_entropy_0.00003_lr_2e-6": ("sequence_mean", 16),
    "factorial_sequence_mean_g32_entropy_0.00003_lr_2e-6": ("sequence_mean", 32),
    "factorial_constdenom_g8_entropy_0.00003_lr_2e-6": ("constant_denominator", 8),
    "factorial_constdenom_g16_entropy_0.00003_lr_2e-6": ("constant_denominator", 16),
    "factorial_constdenom_g32_entropy_0.00003_lr_2e-6": ("constant_denominator", 32),
    "factorial_constdenom_g64_entropy_0.00003_lr_2e-6": ("constant_denominator", 64),
}

GROUP_COLORS = {4: "#E69F00", 8: "#0072B2", 16: "#D55E00", 32: "#009E73", 64: "#CC79A7"}
LOSS_STYLES = {"sequence_mean": "--", "constant_denominator": "-"}
# Non-paired Qwen3-Coder archive estimates published from marin issue #8809:
# https://github.com/marin-community/marin/issues/8809#issuecomment-5516287195
BASE_REFERENCES = {
    "BugsInPy": {"pass@1": 0.3049, "pass@16": 0.5761},
    "SWE-Gym": {"pass@1": 0.3943, "pass@16": 0.9084},
}
BASE_STYLES = {"pass@1": ":", "pass@16": "-."}


def main() -> None:
    frame = pd.read_csv(INPUT)
    frame = frame[
        frame["status"].isin(["authoritative", "authoritative_policy_stop", "authoritative_user_stop"])
    ].copy()
    frame = frame[frame["recipe"].isin(SERIES)].copy()

    # A restarted arm can have more than one authoritative attempt. Retain only
    # the latest attempt so disconnected histories are not presented as a curve.
    latest_attempts = frame.groupby(["dataset", "recipe"])["attempt"].transform("max")
    frame = frame[frame["attempt"] == latest_attempts]

    peaks = frame.groupby(["dataset", "recipe"])["pass_at_1"].max()
    leaders = peaks.groupby(level="dataset").idxmax().to_dict()

    plt.rcParams.update(
        {
            "font.family": "DejaVu Sans",
            "font.size": 12,
            "axes.titlesize": 17,
            "axes.titleweight": "bold",
            "axes.labelsize": 13,
            "legend.fontsize": 10.5,
        }
    )
    fig, axes = plt.subplots(1, 2, figsize=(15.5, 8.5), sharey=True)

    for ax, dataset in zip(axes, ("BugsInPy", "SWE-Gym"), strict=True):
        panel = frame[frame["dataset"] == dataset]
        for recipe, (loss_reduction, group_size) in SERIES.items():
            series = panel[panel["recipe"] == recipe].sort_values("step")
            if series.empty:
                continue
            is_leader = leaders[dataset][1] == recipe
            ax.plot(
                series["step"],
                series["pass_at_1"],
                color=GROUP_COLORS[group_size],
                linestyle=LOSS_STYLES[loss_reduction],
                marker="o",
                markersize=6 if is_leader else 5,
                linewidth=3.0 if is_leader else 1.8,
                alpha=1.0 if is_leader else 0.72,
                zorder=3 if is_leader else 2,
            )
            peak_row = series.loc[series["pass_at_1"].idxmax()]
            ax.scatter(
                [peak_row["step"]],
                [peak_row["pass_at_1"]],
                marker="D",
                s=68 if is_leader else 48,
                color=GROUP_COLORS[group_size],
                edgecolor="white",
                linewidth=0.9,
                zorder=4,
            )
            stop_rows = series[series["status"].isin(["authoritative_policy_stop", "authoritative_user_stop"])]
            if not stop_rows.empty:
                stop_row = stop_rows.sort_values("step").iloc[-1]
                ax.scatter(
                    [stop_row["step"]],
                    [stop_row["pass_at_1"]],
                    marker="X",
                    s=105 if is_leader else 82,
                    color=GROUP_COLORS[group_size],
                    edgecolor="white",
                    linewidth=1.0,
                    zorder=5,
                )

        for metric, value in BASE_REFERENCES[dataset].items():
            ax.axhline(
                value,
                color="#555555",
                linestyle=BASE_STYLES[metric],
                linewidth=1.6,
                alpha=0.82,
                zorder=1,
            )
            ax.text(
                0.985,
                value + 0.008,
                f"Archive base {metric} {value:.4f}",
                transform=ax.get_yaxis_transform(),
                ha="right",
                va="bottom",
                fontsize=9.5,
                color="#444444",
                bbox={"facecolor": "white", "edgecolor": "none", "alpha": 0.82, "pad": 1.4},
                zorder=6,
            )

        ax.set_title(dataset, pad=10)
        ax.set_xlabel("Training step")
        ax.set_xlim(-0.35, max(8.8, float(panel["step"].max()) + 0.8))
        ax.set_ylim(0, 1.0)
        ax.set_xticks(sorted(panel["step"].unique()))
        ax.grid(True, color="#CFD5DC", linewidth=0.7, alpha=0.65)
        ax.set_axisbelow(True)

    axes[0].set_ylabel("Fixed-validation pass@1")
    fig.suptitle("Software-transfer performance so far", fontsize=21, fontweight="bold", y=0.96)
    fig.text(
        0.5,
        0.905,
        "Seed-42 disjoint 128-task validation; matched factorial arms only. Diamonds mark peaks; X marks termination; bold lines lead.\n"
        "Gray lines are non-paired Qwen3-Coder archive references from Marin issue #8809.",
        ha="center",
        va="top",
        fontsize=12.5,
        color="#444444",
    )
    legend_handles = [
        *[
            Line2D(
                [0],
                [0],
                color=GROUP_COLORS[group_size],
                marker="o",
                linestyle="none",
                markersize=7,
                label=f"Group {group_size}",
            )
            for group_size in GROUP_COLORS
        ],
        Line2D([0], [0], color="#333333", linestyle="--", linewidth=2.2, label="Sequence mean"),
        Line2D([0], [0], color="#333333", linestyle="-", linewidth=2.2, label="Constant denominator"),
        Line2D(
            [0],
            [0],
            color="#555555",
            marker="X",
            markeredgecolor="white",
            linestyle="none",
            markersize=9,
            label="Terminated",
        ),
        Line2D(
            [0],
            [0],
            color="#555555",
            linestyle=BASE_STYLES["pass@1"],
            linewidth=1.8,
            label="Archive base pass@1",
        ),
        Line2D(
            [0],
            [0],
            color="#555555",
            linestyle=BASE_STYLES["pass@16"],
            linewidth=1.8,
            label="Archive base pass@16",
        ),
    ]
    fig.legend(
        handles=legend_handles,
        loc="lower center",
        bbox_to_anchor=(0.5, 0.025),
        ncol=5,
        frameon=False,
        title="Color = group size; line = loss reduction",
    )
    fig.tight_layout(rect=(0, 0.18, 1, 0.84), w_pad=3.0)
    fig.savefig(OUTPUT.with_suffix(".png"), dpi=180, bbox_inches="tight")
    fig.savefig(OUTPUT.with_suffix(".svg"), bbox_inches="tight")


if __name__ == "__main__":
    main()