File size: 6,403 Bytes
a2144da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import argparse
import json
from pathlib import Path
from typing import Any

import matplotlib.pyplot as plt


SERIES_ORDER = [
    ("center_sft", "center SFT", "#2563eb"),
    ("center_grpo", "center GRPO", "#7c3aed"),
    ("warehouse_sft", "warehouse SFT", "#0f766e"),
    ("warehouse_grpo", "warehouse GRPO", "#047857"),
    ("center", "center", "#2563eb"),
    ("warehouse", "warehouse", "#0f766e"),
]


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Create static SupplyMind training/eval plots.")
    parser.add_argument("--input", type=Path, default=Path("results/training_dashboard.json"))
    parser.add_argument("--output-dir", type=Path, default=Path("results/plots"))
    return parser.parse_args()


def load_json(path: Path) -> dict[str, Any]:
    return json.loads(path.read_text(encoding="utf-8"))


def value(row: dict[str, Any], *keys: str) -> float | None:
    for key in keys:
        raw = row.get(key)
        if raw is None:
            continue
        try:
            return float(raw)
        except (TypeError, ValueError):
            continue
    return None


def active_series(data: dict[str, Any]) -> list[tuple[str, str, str]]:
    series = data.get("training_series", {})
    active = [item for item in SERIES_ORDER if item[0] in series]
    if "center_sft" in series or "center_grpo" in series or "warehouse_sft" in series or "warehouse_grpo" in series:
        return [item for item in active if item[0].endswith(("_sft", "_grpo"))]
    return active


def line_plot(data: dict[str, Any], y_keys: tuple[str, ...], title: str, ylabel: str, output: Path) -> None:
    series = data.get("training_series", {})
    plt.figure(figsize=(10, 5.2))
    plotted = False
    for key, label, color in active_series(data):
        rows = series.get(key, {}).get("steps", [])
        xs: list[float] = []
        ys: list[float] = []
        for idx, row in enumerate(rows, start=1):
            y = value(row, *y_keys)
            if y is None:
                continue
            xs.append(value(row, "step", "global_step") or idx)
            ys.append(y)
        if xs:
            plt.plot(xs, ys, label=label, color=color, linewidth=2)
            plotted = True
    plt.title(title)
    plt.xlabel("training step")
    plt.ylabel(ylabel)
    plt.grid(alpha=0.25)
    if plotted:
        plt.legend()
    else:
        plt.text(0.5, 0.5, "No series available", ha="center", va="center", transform=plt.gca().transAxes)
    plt.tight_layout()
    plt.savefig(output, dpi=160)
    plt.close()


def invalid_plot(data: dict[str, Any], output: Path) -> None:
    series = data.get("training_series", {})
    labels: list[str] = []
    payloads: list[float] = []
    actions: list[float] = []
    for key, label, _color in active_series(data):
        batches = series.get(key, {}).get("reward_batches", [])
        if not batches:
            continue
        labels.append(label)
        payloads.append(sum(value(row, "invalid_payloads") or 0 for row in batches))
        actions.append(sum(value(row, "invalid_actions") or 0 for row in batches))
    plt.figure(figsize=(10, 5.2))
    if labels:
        xs = range(len(labels))
        plt.bar([x - 0.18 for x in xs], payloads, width=0.36, label="invalid payloads", color="#c2410c")
        plt.bar([x + 0.18 for x in xs], actions, width=0.36, label="invalid env actions", color="#b7791f")
        plt.xticks(list(xs), labels, rotation=20, ha="right")
        plt.legend()
    else:
        plt.text(0.5, 0.5, "No invalid-action diagnostics available", ha="center", va="center", transform=plt.gca().transAxes)
    plt.title("Invalid Payloads / Actions")
    plt.ylabel("count across logged reward batches")
    plt.grid(axis="y", alpha=0.25)
    plt.tight_layout()
    plt.savefig(output, dpi=160)
    plt.close()


def heldout_plot(data: dict[str, Any], output: Path) -> None:
    comparisons = data.get("comparisons", [])
    roles = ["center", "warehouse"]
    variants = [("base", "#64748b"), ("sft", "#2563eb"), ("grpo", "#7c3aed")]
    groups: list[tuple[str, str, str]] = []
    for role in roles:
        groups.append((role, "global", "mean_global_score"))
        groups.append((role, "role", "mean_center_role_score" if role == "center" else "mean_warehouse_role_score"))

    plt.figure(figsize=(11, 5.5))
    group_xs = list(range(len(groups)))
    width = 0.22
    any_rows = False
    for offset, (variant, color) in zip([-width, 0, width], variants, strict=True):
        ys = []
        for role, _metric, key in groups:
            row = next((item for item in comparisons if item.get("role") == role and item.get("label") == variant), None)
            ys.append(value(row or {}, key) if row else None)
        xs = [x + offset for x, y in zip(group_xs, ys, strict=True) if y is not None]
        vals = [y for y in ys if y is not None]
        if vals:
            any_rows = True
            plt.bar(xs, vals, width=width, label=variant.upper(), color=color)

    if any_rows:
        plt.xticks(group_xs, [f"{role}\n{metric}" for role, metric, _key in groups])
        plt.ylim(0, 1)
        plt.legend()
    else:
        plt.text(0.5, 0.5, "No held-out comparisons available", ha="center", va="center", transform=plt.gca().transAxes)
    plt.title("Held-out Scores: Base vs SFT vs GRPO")
    plt.ylabel("normalized score")
    plt.grid(axis="y", alpha=0.25)
    plt.tight_layout()
    plt.savefig(output, dpi=160)
    plt.close()


def main() -> None:
    args = parse_args()
    data = load_json(args.input)
    args.output_dir.mkdir(parents=True, exist_ok=True)
    line_plot(data, ("loss",), "Loss Over Step", "loss", args.output_dir / "loss.png")
    line_plot(data, ("reward", "rewards/reward_completions/mean"), "Reward Over Step", "reward", args.output_dir / "reward.png")
    line_plot(data, ("completions/clipped_ratio", "clipped_ratio"), "Clipped Ratio Over Step", "clipped ratio", args.output_dir / "clipped_ratio.png")
    line_plot(data, ("completions/mean_length", "completion_length", "mean_completion_length"), "Completion Length Over Step", "tokens", args.output_dir / "completion_length.png")
    invalid_plot(data, args.output_dir / "invalids.png")
    heldout_plot(data, args.output_dir / "heldout_comparison.png")
    print(f"Wrote plots to {args.output_dir}")


if __name__ == "__main__":
    main()