File size: 7,645 Bytes
51e7a80
93a239c
 
51e7a80
 
 
93a239c
 
51e7a80
93a239c
51e7a80
 
 
 
93a239c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d0eb59c
51e7a80
93a239c
 
 
51e7a80
93a239c
 
 
51e7a80
 
93a239c
 
 
51e7a80
93a239c
 
 
d0eb59c
 
93a239c
 
 
 
 
 
51e7a80
93a239c
 
 
51e7a80
93a239c
 
 
51e7a80
 
 
 
 
 
93a239c
 
d0eb59c
 
93a239c
 
 
 
51e7a80
 
93a239c
 
d0eb59c
93a239c
 
51e7a80
 
 
 
93a239c
 
 
 
51e7a80
 
93a239c
 
51e7a80
93a239c
 
51e7a80
 
 
 
d0eb59c
93a239c
 
d0eb59c
93a239c
 
 
 
51e7a80
 
 
 
 
 
 
93a239c
 
 
51e7a80
93a239c
 
 
 
 
 
 
 
 
 
 
51e7a80
 
 
 
93a239c
 
 
 
 
 
51e7a80
 
93a239c
 
d0eb59c
 
93a239c
51e7a80
 
 
 
93a239c
 
 
 
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
"""Combined reward curves — Qwen 1.5B vs Llama 1B vs Llama 3B.

Reads:
  --qwen-log     runs/qwen_v6/training_log.jsonl
  --llama-log    runs/llama_v3/training_log.jsonl
  --llama3b-log  runs/llama3b_v1/training_log.jsonl   (optional)

Writes:
  --output-png   eval/plots/reward_curves_combined.png

The headline plot for the rubric: shows three architectures trained on identical
GRPO config. Qwen climbed past zero (after EOS-list fix), Llama-1B plateaued
deeply negative (capacity ceiling), and Llama-3B climbed cleanly without needing
the EOS fix.
"""

from __future__ import annotations

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

import matplotlib

matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np


def _read_jsonl(path: Path) -> list[dict[str, Any]]:
    rows: list[dict[str, Any]] = []
    with path.open() as f:
        for line in f:
            line = line.strip()
            if not line:
                continue
            rows.append(json.loads(line))
    return rows


def _extract_series(rows: list[dict[str, Any]], key: str) -> tuple[np.ndarray, np.ndarray]:
    xs, ys = [], []
    for r in rows:
        v = r.get(key)
        if v is None or (isinstance(v, float) and math.isnan(v)):
            continue
        xs.append(int(r.get("step", r.get("episode", 0))))
        ys.append(float(v))
    return np.array(xs, dtype=np.float64), np.array(ys, dtype=np.float64)


def _smooth(y: np.ndarray, window: int = 3) -> np.ndarray:
    if window <= 1 or len(y) < window:
        return y.copy()
    kernel = np.ones(window, dtype=np.float64) / window
    pad = window - 1
    yp = np.concatenate([np.full(pad, y[0]), y])
    return np.convolve(yp, kernel, mode="valid")


def plot_combined(
    qwen_log: Path,
    llama_log: Path,
    output_png: Path,
    smooth_window: int = 3,
    xkcd: bool = False,
    llama3b_log: Path | None = None,
) -> None:
    qwen_rows = _read_jsonl(qwen_log)
    llama_rows = _read_jsonl(llama_log)
    llama3b_rows = _read_jsonl(llama3b_log) if llama3b_log and llama3b_log.exists() else []

    qx, qy = _extract_series(qwen_rows, "reward")
    lx, ly = _extract_series(llama_rows, "reward")
    l3x, l3y = (_extract_series(llama3b_rows, "reward")
                if llama3b_rows else (np.array([]), np.array([])))

    qy_smooth = _smooth(qy, smooth_window)
    ly_smooth = _smooth(ly, smooth_window)
    l3y_smooth = _smooth(l3y, smooth_window) if len(l3y) > 0 else l3y

    output_png.parent.mkdir(parents=True, exist_ok=True)

    if xkcd:
        plt.xkcd(scale=1.0, length=100, randomness=2)
    fig, (ax_main, ax_clip) = plt.subplots(
        2, 1, figsize=(11, 8), dpi=200, gridspec_kw={"height_ratios": [3, 1.2]}, sharex=True
    )

    QWEN_COLOR = "#1f77b4"
    LLAMA_COLOR = "#888888"
    LLAMA3B_COLOR = "#2ca02c"

    ax_main.scatter(qx, qy, s=18, alpha=0.35, color=QWEN_COLOR)
    ax_main.plot(qx, qy_smooth, color=QWEN_COLOR, linewidth=2.4,
                 label=f"Qwen2.5-1.5B (final={qy[-1]:+.3f}, peak={qy.max():+.3f})")

    ax_main.scatter(lx, ly, s=18, alpha=0.35, color=LLAMA_COLOR)
    ax_main.plot(lx, ly_smooth, color=LLAMA_COLOR, linewidth=2.4,
                 label=f"Llama-3.2-1B (final={ly[-1]:+.3f}, peak={ly.max():+.3f})")

    if len(l3y) > 0:
        ax_main.scatter(l3x, l3y, s=18, alpha=0.35, color=LLAMA3B_COLOR)
        ax_main.plot(l3x, l3y_smooth, color=LLAMA3B_COLOR, linewidth=2.4,
                     label=f"Llama-3.2-3B (final={l3y[-1]:+.3f}, peak={l3y.max():+.3f})")

    ax_main.axhline(0.0, color="#000000", linestyle="-", linewidth=0.8, alpha=0.5)
    floor_style = "-" if xkcd else ":"
    ax_main.axhline(-1.0, color="#d62728", linestyle=floor_style, linewidth=1.0, alpha=0.6,
                    label="reward floor (parser fails)")

    ax_main.set_ylabel("Reward (per-step mean across G=4 rollouts)", fontsize=11)
    ax_main.set_title(
        "Viveka GRPO Training — Three Architectures, Identical Config\n"
        "Qwen 1.5B climbed past zero after EOS-list fix; Llama 1B plateaued; Llama 3B climbed cleanly",
        fontsize=12,
    )
    ax_main.grid(True, alpha=0.3, linestyle="-" if xkcd else ":")
    ax_main.legend(loc="lower right", frameon=True, fontsize=10)

    all_mins = [-1.0, qy.min(), ly.min()] + ([l3y.min()] if len(l3y) > 0 else [])
    all_maxes = [0.3, qy.max(), ly.max()] + ([l3y.max()] if len(l3y) > 0 else [])
    y_min = min(all_mins) - 0.05
    y_max = max(all_maxes) + 0.05
    ax_main.set_ylim(y_min, y_max)

    qx_clip, qy_clip = _extract_series(qwen_rows, "clipped_ratio")
    lx_clip, ly_clip = _extract_series(llama_rows, "clipped_ratio")
    l3x_clip, l3y_clip = (_extract_series(llama3b_rows, "clipped_ratio")
                          if llama3b_rows else (np.array([]), np.array([])))
    if len(qy_clip) > 0:
        ax_clip.plot(qx_clip, qy_clip, color=QWEN_COLOR, linewidth=2.0, marker="o", markersize=4,
                     label=f"Qwen 1.5B (final={qy_clip[-1]:.3f})")
    if len(ly_clip) > 0:
        ax_clip.plot(lx_clip, ly_clip, color=LLAMA_COLOR, linewidth=2.0, marker="s", markersize=4,
                     label=f"Llama 1B (final={ly_clip[-1]:.3f})")
    if len(l3y_clip) > 0:
        ax_clip.plot(l3x_clip, l3y_clip, color=LLAMA3B_COLOR, linewidth=2.0, marker="^", markersize=4,
                     label=f"Llama 3B (final={l3y_clip[-1]:.3f})")
    ax_clip.set_ylabel("Clipped ratio\n(lower = healthier)", fontsize=10)
    ax_clip.set_xlabel("Training step", fontsize=11)
    ax_clip.set_ylim(0.0, 1.05)
    ax_clip.grid(True, alpha=0.3, linestyle="-" if xkcd else ":")
    ax_clip.legend(loc="upper right", frameon=True, fontsize=9)

    qy_final = float(qy[-1])
    ly_final = float(ly[-1])
    box_lines = [
        f"Qwen 1.5B final:  {qy_final:+.3f}",
        f"Llama 1B final:   {ly_final:+.3f}",
    ]
    if len(l3y) > 0:
        l3y_final = float(l3y[-1])
        box_lines.append(f"Llama 3B final:   {l3y_final:+.3f}")
    ax_main.text(
        0.02,
        0.97,
        "\n".join(box_lines),
        transform=ax_main.transAxes,
        fontsize=10,
        verticalalignment="top",
        family="monospace",
        bbox=dict(boxstyle="round,pad=0.5", facecolor="white", alpha=0.92, edgecolor="#cccccc"),
    )

    fig.tight_layout()
    fig.savefig(output_png, dpi=200, bbox_inches="tight")
    plt.close(fig)
    print(f"wrote {output_png}")
    print(f"  Qwen 1.5B:  n={len(qy)} steps, final={qy_final:+.4f}, peak={qy.max():+.4f}")
    print(f"  Llama 1B:   n={len(ly)} steps, final={ly_final:+.4f}, peak={ly.max():+.4f}")
    if len(l3y) > 0:
        print(f"  Llama 3B:   n={len(l3y)} steps, final={float(l3y[-1]):+.4f}, peak={l3y.max():+.4f}")


def main() -> None:
    p = argparse.ArgumentParser(description=__doc__)
    p.add_argument("--qwen-log", type=Path, default=Path("runs/qwen_v6/training_log.jsonl"))
    p.add_argument("--llama-log", type=Path, default=Path("runs/llama_v3/training_log.jsonl"))
    p.add_argument("--llama3b-log", type=Path, default=Path("runs/llama3b_v1/training_log.jsonl"),
                   help="Llama 3B training log (set to /dev/null to skip)")
    p.add_argument("--output-png", type=Path, default=Path("eval/plots/reward_curves_combined.png"))
    p.add_argument("--smooth-window", type=int, default=3)
    p.add_argument("--xkcd", action="store_true",
                   help="Render in xkcd / hand-drawn style for the README hero image")
    args = p.parse_args()
    plot_combined(
        args.qwen_log, args.llama_log, args.output_png, args.smooth_window,
        args.xkcd, args.llama3b_log,
    )


if __name__ == "__main__":
    main()