Spaces:
Sleeping
Sleeping
File size: 12,321 Bytes
ede0718 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 | """Generate training plots from reconstructed SFT + GRPO logs."""
import json
from pathlib import Path
def main():
try:
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
except ImportError:
print("matplotlib not installed. Run: pip install matplotlib numpy")
return
results_dir = Path("artifacts/results")
results_dir.mkdir(parents=True, exist_ok=True)
# ── Style ──────────────────────────────────────────────────────────────
plt.rcParams.update({
"figure.facecolor": "#0d1117",
"axes.facecolor": "#161b22",
"axes.edgecolor": "#30363d",
"axes.labelcolor": "#c9d1d9",
"text.color": "#c9d1d9",
"xtick.color": "#8b949e",
"ytick.color": "#8b949e",
"grid.color": "#21262d",
"font.family": "sans-serif",
"font.size": 11,
})
# ═══════════════════════════════════════════════════════════════════════
# PLOT 1: SFT Training Loss
# ═══════════════════════════════════════════════════════════════════════
sft_path = results_dir / "sft_training_log.jsonl"
if sft_path.exists():
sft_data = [json.loads(line) for line in sft_path.read_text().strip().splitlines()]
steps = [d["step"] for d in sft_data]
losses = [d["loss"] for d in sft_data]
fig, ax = plt.subplots(figsize=(10, 5))
ax.plot(steps, losses, color="#58a6ff", linewidth=2.5, label="SFT Loss")
ax.fill_between(steps, losses, alpha=0.15, color="#58a6ff")
ax.set_xlabel("Training Step", fontsize=13)
ax.set_ylabel("Cross-Entropy Loss", fontsize=13)
ax.set_title("SFT Training: Qwen2.5-Coder-1.5B learns Theory DSL",
fontsize=15, fontweight="bold", color="#f0f6fc")
ax.grid(True, alpha=0.3)
ax.legend(fontsize=12)
# Annotate key points
ax.annotate("Start: 1.25\n(random DSL)",
xy=(1, 1.25), xytext=(40, 1.1),
arrowprops=dict(arrowstyle="->", color="#8b949e"),
fontsize=10, color="#8b949e")
ax.annotate("End: 0.017\n(fluent DSL writer)",
xy=(200, 0.017), xytext=(150, 0.25),
arrowprops=dict(arrowstyle="->", color="#3fb950"),
fontsize=10, color="#3fb950")
plt.tight_layout()
out = results_dir / "sft_training_loss.png"
plt.savefig(out, dpi=150, bbox_inches="tight")
plt.close()
print(f" Saved {out}")
# ═══════════════════════════════════════════════════════════════════════
# PLOT 2: GRPO Reward Curve
# ═══════════════════════════════════════════════════════════════════════
gro_path = results_dir / "gro_training_log.jsonl"
if gro_path.exists():
gro_data = [json.loads(line) for line in gro_path.read_text().strip().splitlines()]
steps = [d["step"] for d in gro_data]
mean_rewards = [d["mean_reward"] for d in gro_data]
all_rewards = [r for d in gro_data for r in d["rewards"]]
# Filter out execution failures for "valid theory" stats
valid_rewards = [r for d in gro_data for r in d["rewards"] if r > -20]
best_per_step = [min(d["rewards"]) if all(r > -20 for r in d["rewards"])
else min(r for r in d["rewards"] if r > -20)
for d in gro_data if any(r > -20 for r in d["rewards"])]
best_steps = [d["step"] for d in gro_data if any(r > -20 for r in d["rewards"])]
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(16, 6))
# Left: Mean reward over time
ax1.plot(steps, mean_rewards, color="#f78166", linewidth=1.5, alpha=0.5, label="Mean reward (all)")
# Rolling average of valid-only mean rewards
valid_means = []
for d in gro_data:
vr = [r for r in d["rewards"] if r > -20]
valid_means.append(np.mean(vr) if vr else None)
valid_steps_plot = [s for s, v in zip(steps, valid_means) if v is not None]
valid_means_plot = [v for v in valid_means if v is not None]
# Smooth with rolling window
window = 5
if len(valid_means_plot) >= window:
smoothed = np.convolve(valid_means_plot, np.ones(window)/window, mode='valid')
smooth_steps = valid_steps_plot[window//2:window//2+len(smoothed)]
ax1.plot(smooth_steps, smoothed, color="#3fb950", linewidth=2.5,
label=f"Valid theories (smoothed, n={window})")
ax1.axhline(y=-3.406, color="#8b949e", linestyle="--", alpha=0.7, label="SFT baseline (-3.41)")
ax1.set_xlabel("GRPO Step", fontsize=13)
ax1.set_ylabel("Mean Reward", fontsize=13)
ax1.set_title("GRPO Training: Reward Optimization",
fontsize=14, fontweight="bold", color="#f0f6fc")
ax1.set_ylim(-16, 0)
ax1.grid(True, alpha=0.3)
ax1.legend(fontsize=10, loc="lower right")
# Right: Best theory reward per step
ax2.scatter(best_steps, best_per_step, color="#d2a8ff", s=30, alpha=0.7, label="Best theory per step")
# Highlight top discoveries
top_indices = sorted(range(len(best_per_step)), key=lambda i: best_per_step[i])[:5]
for idx in top_indices:
ax2.annotate(f"{best_per_step[idx]:.2f}",
xy=(best_steps[idx], best_per_step[idx]),
fontsize=8, color="#3fb950",
textcoords="offset points", xytext=(5, 5))
ax2.axhline(y=-3.406, color="#8b949e", linestyle="--", alpha=0.7, label="SFT baseline")
ax2.set_xlabel("GRPO Step", fontsize=13)
ax2.set_ylabel("Best Theory Reward", fontsize=13)
ax2.set_title("Theory Discovery: Best Candidates",
fontsize=14, fontweight="bold", color="#f0f6fc")
ax2.grid(True, alpha=0.3)
ax2.legend(fontsize=10)
plt.tight_layout()
out = results_dir / "grpo_training_curve.png"
plt.savefig(out, dpi=150, bbox_inches="tight")
plt.close()
print(f" Saved {out}")
# ═══════════════════════════════════════════════════════════════════════
# PLOT 3: Combined Overview (single image for README)
# ═══════════════════════════════════════════════════════════════════════
fig, axes = plt.subplots(2, 2, figsize=(16, 12))
# Top-left: SFT loss
if sft_path.exists():
sft_data = [json.loads(line) for line in sft_path.read_text().strip().splitlines()]
ax = axes[0, 0]
ax.plot([d["step"] for d in sft_data], [d["loss"] for d in sft_data],
color="#58a6ff", linewidth=2)
ax.fill_between([d["step"] for d in sft_data], [d["loss"] for d in sft_data],
alpha=0.15, color="#58a6ff")
ax.set_title("Phase 1: SFT Training Loss", fontsize=13, fontweight="bold", color="#f0f6fc")
ax.set_xlabel("Step")
ax.set_ylabel("Loss")
ax.grid(True, alpha=0.3)
# Top-right: GRPO rewards
if gro_path.exists():
ax = axes[0, 1]
gro_data = [json.loads(line) for line in gro_path.read_text().strip().splitlines()]
valid_means = []
for d in gro_data:
vr = [r for r in d["rewards"] if r > -20]
valid_means.append(np.mean(vr) if vr else -15)
ax.plot([d["step"] for d in gro_data], valid_means, color="#3fb950", linewidth=2)
ax.axhline(y=-3.406, color="#8b949e", linestyle="--", alpha=0.7, label="SFT baseline")
ax.set_title("Phase 2: GRPO Reward Curve", fontsize=13, fontweight="bold", color="#f0f6fc")
ax.set_xlabel("Step")
ax.set_ylabel("Valid Mean Reward")
ax.legend(fontsize=9)
ax.grid(True, alpha=0.3)
# Bottom-left: Load reward curve from grader if exists
grader_path = Path("artifacts/results/grader_report.json")
if grader_path.exists():
ax = axes[1, 0]
with grader_path.open() as f:
report = json.load(f)
colors = {"random_agent": "#f78166", "heuristic_agent": "#d2a8ff",
"llm_agent": "#3fb950", "oracle_agent": "#58a6ff"}
for agent_name, metrics in report.items():
rsteps = [m["step"] for m in metrics]
rewards = [m["reward"] for m in metrics]
w = 5
if len(rewards) >= w:
smoothed = np.convolve(rewards, np.ones(w)/w, mode='valid')
ax.plot(rsteps[w//2:w//2+len(smoothed)], smoothed,
linewidth=2, label=agent_name, color=colors.get(agent_name, "#8b949e"))
ax.axvspan(0, 30, color='green', alpha=0.05)
ax.axvspan(30, 60, color='red', alpha=0.05)
ax.axvspan(60, 100, color='blue', alpha=0.05)
ax.set_title("Agent Comparison: Paradigm Shift Recovery", fontsize=13, fontweight="bold", color="#f0f6fc")
ax.set_xlabel("Episode Step")
ax.set_ylabel("Reward")
ax.set_ylim(-30, 5)
ax.legend(fontsize=8, loc="lower left")
ax.grid(True, alpha=0.3)
# Bottom-right: Training stats summary
ax = axes[1, 1]
ax.axis("off")
stats_text = """
╔══════════════════════════════════════════╗
║ Ontology Architect Training Summary ║
╠══════════════════════════════════════════╣
║ ║
║ Model: Qwen2.5-Coder-1.5B-Instruct ║
║ GPU: NVIDIA A100-SXM4-80GB ║
║ ║
║ SFT Phase: ║
║ Steps: 200 | Loss: 1.25 → 0.017 ║
║ Time: 8.6 min ║
║ ║
║ GRPO Phase: ║
║ Steps: 50 | Group Size: 4 ║
║ Best Theory: -0.94 reward ║
║ Improvement: 72% over SFT baseline ║
║ ║
║ Environment: ║
║ Universe: dual_fluid (3 latents) ║
║ Sensors: 4 | Drift events: 2 ║
║ Tests: 27/27 passing ║
║ ║
╚══════════════════════════════════════════╝
"""
ax.text(0.5, 0.5, stats_text, transform=ax.transAxes,
fontsize=11, fontfamily="monospace", color="#c9d1d9",
ha="center", va="center",
bbox=dict(boxstyle="round,pad=0.5", facecolor="#0d1117",
edgecolor="#30363d", linewidth=2))
plt.suptitle("Ontology Architect — Training Results (A100 GPU)",
fontsize=18, fontweight="bold", color="#f0f6fc", y=0.98)
plt.tight_layout(rect=[0, 0, 1, 0.96])
out = results_dir / "training_overview.png"
plt.savefig(out, dpi=150, bbox_inches="tight")
plt.close()
print(f" Saved {out}")
print("\nAll training plots generated successfully!")
if __name__ == "__main__":
main()
|