File size: 17,120 Bytes
35e9f20 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 | """
plot_results.py β Results Charts for Hackathon Presentation
============================================================
Generates publication-ready charts from evaluation JSON data.
Saves all charts to results/charts/ as both PNG and SVG.
BASELINE CONSTANTS (FINAL β measured, do not change):
R1 Llama-3.1-8B zero-shot: easy=0.0100, medium=0.4583, hard=0.0100, avg=0.1594
R2 Llama-3.1-8B zero-shot: easy=0.3198, medium=0.2443, hard=0.2520, avg=0.2720
Training model: Qwen/Qwen2.5-1.5B-Instruct (GRPO, 4-bit QLoRA)
Charts produced:
1. r1_scores_comparison.png β R1 Llama baseline vs trained bar chart
2. r2_scores_comparison.png β R2 Llama zero-shot vs trained bar chart (no rule-based)
3. sprint_rewards.png β Sprint-by-sprint reward for each R2 scenario
4. improvement_summary.png β Combined before/after delta chart (main slide chart)
5. training_curve.png β GRPO training loss/reward curve (if trainer_state.json present)
Usage:
# After running evaluate_r2.py --baseline-only:
python plot_results.py --eval results/r2_evaluation.json
# With training curve (after train_llm.py):
python plot_results.py --eval results/r2_evaluation.json \\
--trainer results/trained_model/trainer_state.json
# Hardcode known scores for presentation (no eval file needed):
python plot_results.py --demo
"""
from __future__ import annotations
import argparse
import json
import os
from pathlib import Path
CHARTS_DIR = Path("results/charts")
CHARTS_DIR.mkdir(parents=True, exist_ok=True)
# ββ Measured baselines (FINAL) ββββββββββββββββββββββββββββββββββββββββββββββββ
LLAMA_BASELINE_R1 = {
"easy_sprint": 0.0100,
"medium_sprint": 0.4583,
"hard_sprint": 0.0100,
"average": 0.1594,
}
LLAMA_BASELINE_R2 = {
"project_easy": 0.3198,
"project_medium": 0.2443,
"project_hard": 0.2520,
"average": 0.2720,
}
TRAINING_MODEL = "Qwen/Qwen2.5-1.5B-Instruct"
# ββ Colour palette ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
C_LLAMA = "#6B7280" # grey β Llama zero-shot baseline
C_RULE = "#3B82F6" # blue β rule-based baseline
C_TRAINED = "#10B981" # green β trained Qwen (post-GRPO)
C_EASY = "#60A5FA"
C_MEDIUM = "#F59E0B"
C_HARD = "#EF4444"
C_BG = "#F9FAFB"
C_TEXT = "#111827"
def _setup_matplotlib():
import matplotlib
matplotlib.use("Agg") # headless
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
plt.rcParams.update({
"figure.facecolor": C_BG,
"axes.facecolor": C_BG,
"axes.edgecolor": "#D1D5DB",
"axes.labelcolor": C_TEXT,
"text.color": C_TEXT,
"xtick.color": C_TEXT,
"ytick.color": C_TEXT,
"grid.color": "#E5E7EB",
"grid.linestyle": "--",
"grid.alpha": 0.7,
"font.family": "sans-serif",
"font.size": 11,
"axes.titlesize": 13,
"axes.titleweight": "bold",
"figure.dpi": 150,
})
return plt, mpatches
def save(plt, name: str):
png = CHARTS_DIR / f"{name}.png"
svg = CHARTS_DIR / f"{name}.svg"
plt.tight_layout()
plt.savefig(png, bbox_inches="tight")
plt.savefig(svg, bbox_inches="tight")
plt.close()
print(f" [SAVED] {png}", flush=True)
# ββ Chart 1: R1 scores comparison ββββββββββββββββββββββββββββββββββββββββββββ
def chart_r1_comparison(eval_data: dict):
plt, mpatches = _setup_matplotlib()
tasks = ["easy_sprint", "medium_sprint", "hard_sprint"]
labels = ["Easy Sprint", "Medium Sprint", "Hard Sprint"]
llama_base = [eval_data.get("r1_llama_baseline", LLAMA_BASELINE_R1).get(t, 0) for t in tasks]
rule_based = [eval_data.get("r1_rule_based", {}).get(t, {}).get("avg_score", 0) for t in tasks]
llm_scores = [eval_data.get("r1_llm", {}).get(t, {}).get("avg_score", 0) for t in tasks]
has_llm = any(v > 0 for v in llm_scores)
x = range(len(tasks))
width = 0.28 if has_llm else 0.38
fig, ax = plt.subplots(figsize=(9, 5))
b1 = ax.bar([i - width for i in x], llama_base, width, label=f"Llama-3.1-8B (zero-shot baseline)", color=C_LLAMA, zorder=3)
b2 = ax.bar([i for i in x], rule_based, width, label="Rule-based", color=C_RULE, zorder=3)
if has_llm:
b3 = ax.bar([i + width for i in x], llm_scores, width, label=f"{TRAINING_MODEL} (GRPO trained)", color=C_TRAINED, zorder=3)
def label_bars(bars):
for bar in bars:
h = bar.get_height()
if h > 0.01:
ax.text(bar.get_x() + bar.get_width()/2, h + 0.01,
f"{h:.2f}", ha="center", va="bottom", fontsize=9)
label_bars(b1); label_bars(b2)
if has_llm: label_bars(b3)
ax.set_xticks(list(x))
ax.set_xticklabels(labels)
ax.set_ylim(0, 1.15)
ax.set_ylabel("Score (0.01 β 0.99)")
ax.set_title("Round 1 β Score Comparison")
ax.legend(loc="upper right")
ax.yaxis.grid(True, zorder=0)
ax.set_axisbelow(True)
save(plt, "r1_scores_comparison")
# ββ Chart 2: R2 scores comparison ββββββββββββββββββββββββββββββββββββββββββββ
def chart_r2_comparison(eval_data: dict):
plt, mpatches = _setup_matplotlib()
tasks = ["project_easy", "project_medium", "project_hard"]
labels = ["Easy (6 sprints)", "Medium (6 sprints)", "Hard (6 sprints)"]
llama_base = [eval_data.get("r2_llama_baseline", LLAMA_BASELINE_R2).get(t, 0) for t in tasks]
llm_scores = [eval_data.get("r2_llm", {}).get(t, {}).get("avg_score", 0) for t in tasks]
has_llm = any(v > 0 for v in llm_scores)
x = range(len(tasks))
width = 0.32 if has_llm else 0.5
fig, ax = plt.subplots(figsize=(9, 5))
b1 = ax.bar([i - width/2 for i in x], llama_base, width, label="Llama-3.1-8B (zero-shot)", color=C_LLAMA, zorder=3)
if has_llm:
b2 = ax.bar([i + width/2 for i in x], llm_scores, width, label=f"{TRAINING_MODEL} (GRPO)", color=C_TRAINED, zorder=3)
for bars in ([b1] + ([b2] if has_llm else [])):
for bar in bars:
h = bar.get_height()
if h > 0.01:
ax.text(bar.get_x() + bar.get_width()/2, h + 0.01,
f"{h:.2f}", ha="center", va="bottom", fontsize=9)
ax.set_xticks(list(x))
ax.set_xticklabels(labels)
ax.set_ylim(0, 1.15)
ax.set_ylabel("Project Score (delivery Γ instruction Γ health)")
ax.set_title("Round 2 β Multi-Sprint Project Score")
ax.legend(loc="upper right")
ax.yaxis.grid(True, zorder=0)
ax.set_axisbelow(True)
save(plt, "r2_scores_comparison")
# ββ Chart 3: Sprint reward curves βββββββββββββββββββββββββββββββββββββββββββββ
def chart_sprint_rewards(eval_data: dict):
"""Per-sprint reward for each R2 scenario (Llama baseline vs trained)."""
plt, _ = _setup_matplotlib()
tasks = ["project_easy", "project_medium", "project_hard"]
colors = [C_EASY, C_MEDIUM, C_HARD]
labels = ["Easy", "Medium", "Hard"]
fig, axes = plt.subplots(1, 3, figsize=(13, 4), sharey=True)
for ax, task, color, label in zip(axes, tasks, colors, labels):
llm_eps = eval_data.get("r2_llm", {}).get(task, {}).get("episodes", [])
if llm_eps:
sr = llm_eps[0].get("sprint_rewards", [])
if sr:
ax.plot(range(1, len(sr)+1), sr, "o-",
color=color, label="Qwen GRPO", linewidth=2, markersize=6)
ax.set_xlabel("Sprint")
ax.set_title(f"{label} Project")
ax.set_xticks(range(1, 7))
ax.set_ylim(0, 2.2)
ax.yaxis.grid(True)
ax.set_axisbelow(True)
if ax == axes[0]:
ax.set_ylabel("Sprint Reward")
ax.legend(fontsize=8)
fig.suptitle("Sprint-by-Sprint Reward β Qwen GRPO Trained", y=1.02, fontsize=13, fontweight="bold")
save(plt, "sprint_rewards")
# ββ Chart 4: Improvement summary (main presentation slide) βββββββββββββββββββ
def chart_improvement_summary(eval_data: dict):
"""Main before/after chart. Uses Llama zero-shot as the 'before' bar."""
plt, _ = _setup_matplotlib()
all_tasks = (
[f"R1: {t.replace('_sprint','').title()}" for t in ["easy_sprint","medium_sprint","hard_sprint"]] +
[f"R2: {t.replace('project_','').title()}" for t in ["project_easy","project_medium","project_hard"]]
)
llama_base, trained_scores = [], []
for t in ["easy_sprint","medium_sprint","hard_sprint"]:
llama_base.append(eval_data.get("r1_llama_baseline", LLAMA_BASELINE_R1).get(t, 0))
trained_scores.append(eval_data.get("r1_llm", {}).get(t, {}).get("avg_score", 0))
for t in ["project_easy","project_medium","project_hard"]:
llama_base.append(eval_data.get("r2_llama_baseline", LLAMA_BASELINE_R2).get(t, 0))
trained_scores.append(eval_data.get("r2_llm", {}).get(t, {}).get("avg_score", 0))
x = range(len(all_tasks))
width = 0.36
fig, ax = plt.subplots(figsize=(11, 5))
b1 = ax.bar([i - width/2 for i in x], llama_base, width,
label="Before: Llama-3.1-8B (zero-shot)", color=C_LLAMA, zorder=3)
b2 = ax.bar([i + width/2 for i in x], trained_scores, width,
label=f"After: {TRAINING_MODEL} (GRPO)", color=C_TRAINED, zorder=3)
# Delta arrows and labels
for i, (base, trained) in enumerate(zip(llama_base, trained_scores)):
if trained > base + 0.01:
ax.annotate("", xy=(i + width/2, trained + 0.02), xytext=(i - width/2, base + 0.02),
arrowprops=dict(arrowstyle="->", color="#059669", lw=1.5))
ax.text(i, max(base, trained) + 0.06, f"+{trained-base:.2f}",
ha="center", fontsize=8, color="#059669", fontweight="bold")
ax.set_xticks(list(x))
ax.set_xticklabels(all_tasks, rotation=15, ha="right")
ax.set_ylim(0, 1.25)
ax.set_ylabel("Score")
ax.set_title(f"Reward Improvement: Llama Zero-Shot β Qwen2.5-1.5B GRPO Trained")
ax.legend(loc="upper left")
ax.yaxis.grid(True, zorder=0)
ax.set_axisbelow(True)
# Divider between R1 and R2
ax.axvline(x=2.5, color="#9CA3AF", linestyle=":", linewidth=1.5)
ax.text(1.0, 1.20, "Round 1", ha="center", fontsize=10, color="#6B7280")
ax.text(4.0, 1.20, "Round 2", ha="center", fontsize=10, color="#6B7280")
save(plt, "improvement_summary")
# ββ Chart 5: Training loss/reward curve βββββββββββββββββββββββββββββββββββββββ
def chart_training_curve(trainer_state_path: str):
plt, _ = _setup_matplotlib()
with open(trainer_state_path) as f:
state = json.load(f)
log_history = state.get("log_history", [])
if not log_history:
print(" [SKIP] No log_history in trainer_state.json", flush=True)
return
steps, losses, rewards = [], [], []
for entry in log_history:
if "loss" in entry:
steps.append(entry.get("step", 0))
losses.append(entry["loss"])
if "reward" in entry:
rewards.append((entry.get("step", 0), entry["reward"]))
fig, axes = plt.subplots(1, 2, figsize=(11, 4))
if steps and losses:
axes[0].plot(steps, losses, color=C_RULE, linewidth=2)
axes[0].set_xlabel("Training Step")
axes[0].set_ylabel("Loss")
axes[0].set_title("GRPO Training Loss (Qwen2.5-1.5B)")
axes[0].yaxis.grid(True)
if rewards:
rsteps, rvals = zip(*rewards)
axes[1].plot(rsteps, rvals, color=C_TRAINED, linewidth=2)
axes[1].set_xlabel("Training Step")
axes[1].set_ylabel("Mean Reward")
axes[1].set_title("GRPO Mean Reward per Step")
axes[1].yaxis.grid(True)
fig.suptitle(f"GRPO Training Curves β {TRAINING_MODEL}", fontsize=13, fontweight="bold")
save(plt, "training_curve")
# ββ Demo mode (hardcoded real baselines, placeholder trained scores) ββββββββββ
def demo_mode():
"""
Generate charts using real measured baselines.
Trained scores are placeholders β replace with real evaluate_r2.py output after on-site training.
"""
print("[INFO] Demo mode β real Llama baselines, placeholder trained scores", flush=True)
print(f"[INFO] Training model: {TRAINING_MODEL}", flush=True)
# Placeholder trained scores β update after on-site GRPO training
PLACEHOLDER_R1_TRAINED = {
"easy_sprint": 0.0, # update after training
"medium_sprint": 0.0, # update after training
"hard_sprint": 0.0, # update after training
}
PLACEHOLDER_R2_TRAINED = {
"project_easy": 0.0, # update after training
"project_medium": 0.0, # update after training
"project_hard": 0.0, # update after training
}
demo_data = {
"r1_llama_baseline": LLAMA_BASELINE_R1,
"r2_llama_baseline": LLAMA_BASELINE_R2,
"r1_rule_based": {
"easy_sprint": {"avg_score": 0.92},
"medium_sprint": {"avg_score": 0.35},
"hard_sprint": {"avg_score": 0.01},
},
"r1_llm": {
t: {"avg_score": v} for t, v in PLACEHOLDER_R1_TRAINED.items()
},
"r2_rule_based": {},
"r2_llm": {
t: {"avg_score": v, "episodes": []}
for t, v in PLACEHOLDER_R2_TRAINED.items()
},
"improvement": {
t: {
"llama_baseline": LLAMA_BASELINE_R2[t],
"trained_llm": PLACEHOLDER_R2_TRAINED[t],
"delta_vs_llama": round(PLACEHOLDER_R2_TRAINED[t] - LLAMA_BASELINE_R2[t], 4),
}
for t in ["project_easy", "project_medium", "project_hard"]
},
}
return demo_data
# ββ CLI βββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def main():
parser = argparse.ArgumentParser(description="Generate result charts for presentation")
parser.add_argument("--eval", type=str, default=None,
help="Path to r2_evaluation.json from evaluate_r2.py")
parser.add_argument("--trainer", type=str, default=None,
help="Path to trainer_state.json from train_llm.py output")
parser.add_argument("--demo", action="store_true",
help="Generate charts with real baselines + placeholder trained scores")
args = parser.parse_args()
try:
import matplotlib
except ImportError:
print("[ERROR] matplotlib not installed. Run: pip install matplotlib", flush=True)
import sys; sys.exit(1)
if args.demo:
eval_data = demo_mode()
elif args.eval:
with open(args.eval) as f:
eval_data = json.load(f)
# Back-fill baseline keys if running against old JSON format
if "r1_llama_baseline" not in eval_data:
eval_data["r1_llama_baseline"] = LLAMA_BASELINE_R1
if "r2_llama_baseline" not in eval_data:
eval_data["r2_llama_baseline"] = LLAMA_BASELINE_R2
else:
print("[INFO] No --eval file specified. Using --demo mode.", flush=True)
eval_data = demo_mode()
print(f"\nGenerating charts β {CHARTS_DIR}/", flush=True)
print(f" Baselines: R1 avg={LLAMA_BASELINE_R1['average']:.4f} R2 avg={LLAMA_BASELINE_R2['average']:.4f}", flush=True)
print(f" Training model: {TRAINING_MODEL}", flush=True)
print(" Chart 1: R1 scores comparison...", flush=True)
chart_r1_comparison(eval_data)
print(" Chart 2: R2 scores comparison...", flush=True)
chart_r2_comparison(eval_data)
print(" Chart 3: Sprint reward curves...", flush=True)
chart_sprint_rewards(eval_data)
print(" Chart 4: Improvement summary...", flush=True)
chart_improvement_summary(eval_data)
if args.trainer and Path(args.trainer).exists():
print(" Chart 5: Training curve...", flush=True)
chart_training_curve(args.trainer)
else:
print(" Chart 5: Training curve β skipped (no --trainer file provided)", flush=True)
print(f"\nβ
All charts saved to {CHARTS_DIR}/", flush=True)
print(f" Use improvement_summary.png in your HF blog post and pitch slides.", flush=True)
if __name__ == "__main__":
main() |