Spaces:
Running
Running
File size: 13,536 Bytes
099bec8 84fbeda 099bec8 84fbeda 099bec8 84fbeda 099bec8 aae07d0 84fbeda aae07d0 50d2fcb 84fbeda 50d2fcb 099bec8 | 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 | #!/usr/bin/env python
"""Hackathon-narrative comparison plots that go beyond `make_plots.py`.
Given the eval JSONs that `refresh_all_plots.sh` downloads from each model
repo, this script renders three artefacts targeted at judges:
1. ``06_same_base_delta.png`` β per-family delta (GRPO β base) for each
model size, exposing where RL helps vs. hurts at each scale. This is the
most important hackathon plot: it tells the "scale-dependent training
response" story directly.
2. ``07_runs_summary_table.png`` β clean text table of every run's
aggregate score, format pass rate, and per-family numbers. Ships as a
PNG so it can drop straight into the README.
3. ``runs_summary.json`` β machine-readable version of the same table for
downstream tooling (the blog post inlines it).
Inputs are auto-discovered from ``outputs/run_artifacts/`` so the script
stays in lock-step with whatever has actually been pushed to the Hub by
``refresh_all_plots.sh``. Anything that isn't there yet (e.g. Run 3 / Run
4 evals while training is still in flight) is just omitted from the
matrix β every plot degrades gracefully.
Why this lives outside ``make_plots.py``: ``make_plots.py`` is the
generic per-eval comparison primitive; ``compare_runs.py`` is the
opinionated, run-aware orchestrator that knows the relationship between
the runs (same model size β "delta", different model sizes β side-by-
side).
"""
from __future__ import annotations
import argparse
import json
import statistics
from collections import defaultdict
from dataclasses import dataclass
from pathlib import Path
from typing import Iterable
@dataclass(frozen=True)
class RunSpec:
"""One row of the comparison table.
``base_label`` cross-references the matching base entry by ``label`` so
the same-base delta plot can pair them. ``base_label=None`` means this
row IS itself a base.
"""
label: str
eval_path: Path
base_label: str | None = None
color: str = "tab:blue"
# Ordered for legend stability in plots and rows in the summary table.
RUN_SPECS: list[RunSpec] = [
RunSpec(
label="0.6B base",
eval_path=Path("outputs/run_artifacts/v4/evals/eval_qwen3-0.6b_n50_v4.json"),
color="tab:gray",
),
RunSpec(
label="Probe (0.6B, Ξ²=0)",
eval_path=Path("outputs/run_artifacts/v4/evals/eval_clarify-rl-grpo-qwen3-0-6b_n50_v4.json"),
base_label="0.6B base",
color="tab:blue",
),
RunSpec(
label="1.7B base",
eval_path=Path("outputs/run_artifacts/v4/evals/eval_qwen3-1.7b_n50_v4.json"),
color="dimgray",
),
RunSpec(
label="Drift (1.7B, Ξ²=0)",
eval_path=Path("outputs/run_artifacts/1.7B/evals/eval_clarify-rl-grpo-qwen3-1-7b_n50.json"),
base_label="1.7B base",
color="tab:orange",
),
RunSpec(
label="Anchor (1.7B, Ξ²=0.2)",
# Auto-resolved from outputs/run_artifacts/1.7B-KL/evals/<latest>.json
eval_path=Path("outputs/run_artifacts/1.7B-KL/evals"),
base_label="1.7B base",
color="tab:green",
),
RunSpec(
label="Restrain (1.7B, Ξ²=1.0)",
eval_path=Path("outputs/run_artifacts/1.7B-Run6/evals"),
base_label="1.7B base",
color="#0d47a1",
),
RunSpec(
label="Champion (1.7B, Ξ²=0.3)",
eval_path=Path("outputs/run_artifacts/1.7B-Run7/evals"),
base_label="1.7B base",
color="#ff6f00",
),
RunSpec(
label="4B base",
eval_path=Path("outputs/run_artifacts/4B-base/evals"),
color="darkgray",
),
RunSpec(
label="4B GRPO (Run 3)",
eval_path=Path("outputs/run_artifacts/4B/evals"),
base_label="4B base",
color="tab:purple",
),
RunSpec(
label="4B-instruct",
eval_path=Path("outputs/eval_qwen3-4b-instruct_n50_v4.json"),
color="tab:red",
),
]
def _resolve_eval_path(spec: RunSpec) -> Path | None:
"""If ``spec.eval_path`` is a directory, pick the most-recently-modified
eval JSON inside it. Otherwise return the file as-is. Missing β None.
"""
p = spec.eval_path
if not p.exists():
return None
if p.is_file():
return p
if p.is_dir():
candidates = sorted(
p.glob("eval_*.json"),
key=lambda f: f.stat().st_mtime,
reverse=True,
)
return candidates[0] if candidates else None
return None
def _load_summary_and_results(path: Path) -> tuple[dict, list[dict]]:
data = json.loads(path.read_text())
return data.get("summary", {}), data.get("results", [])
def _per_family_means(results: list[dict]) -> dict[str, float]:
"""Mean final_score per task family. Treats unknown family as ``"?"``."""
by_fam: dict[str, list[float]] = defaultdict(list)
for r in results:
by_fam[r.get("family", "?")].append(float(r.get("final_score", 0.0)))
return {fam: statistics.mean(scores) if scores else 0.0 for fam, scores in by_fam.items()}
def _per_family_max(results: list[dict]) -> dict[str, float]:
by_fam: dict[str, list[float]] = defaultdict(list)
for r in results:
by_fam[r.get("family", "?")].append(float(r.get("final_score", 0.0)))
return {fam: max(scores) if scores else 0.0 for fam, scores in by_fam.items()}
# ---------------------------------------------------------------------------
# Plot 6 β same-base delta chart
# ---------------------------------------------------------------------------
def _all_families(specs: dict[str, dict]) -> list[str]:
fams: set[str] = set()
for entry in specs.values():
fams.update(entry["family_means"].keys())
return sorted(fams)
def _delta_panel(ax, specs: dict[str, dict], pairs, families, metric_key: str, ylabel: str, title: str) -> None:
n_families = len(families)
n_pairs = len(pairs)
width = 0.8 / max(1, n_pairs)
x = list(range(n_families))
for i, (trained_label, entry) in enumerate(pairs):
base_entry = specs[entry["base_label"]]
delta = {
fam: entry[metric_key].get(fam, 0.0) - base_entry[metric_key].get(fam, 0.0)
for fam in families
}
ax.bar(
[xi + (i - (n_pairs - 1) / 2) * width for xi in x],
[delta[fam] for fam in families],
width=width,
label=trained_label,
color=entry["color"],
edgecolor="black",
linewidth=0.5,
)
ax.axhline(0.0, color="black", lw=0.8)
ax.set_xticks(x)
ax.set_xticklabels(families, rotation=15, ha="right")
ax.set_ylabel(ylabel)
ax.set_title(title)
ax.grid(alpha=0.3, axis="y")
def plot_same_base_delta(specs: dict[str, dict], out_path: Path) -> None:
"""For every (trained, base) pair, plot Ξ = trained β base on TWO panels:
left = mean per family (avg behaviour), right = max per family (peak
capability). The right panel exposes the "capability concentration"
finding: Run 2's mean regressed on meeting_scheduling, but its max went
*up*, so it learned a narrower-but-stronger solver for that family.
"""
pairs = [
(label, entry)
for label, entry in specs.items()
if entry["base_label"] and entry["base_label"] in specs
]
if not pairs:
print("[skip] same-base delta β no trained vs base pairs available yet")
return
import matplotlib.pyplot as plt
families = _all_families(specs)
fig, (ax_mean, ax_max) = plt.subplots(1, 2, figsize=(max(13, len(families) * 2.0), 5.5), sharey=False)
_delta_panel(
ax_mean, specs, pairs, families,
metric_key="family_means",
ylabel="Ξ avg score (GRPO β same-size base)",
title="(a) Average behaviour\npositive = GRPO consistently helps, negative = regression",
)
_delta_panel(
ax_max, specs, pairs, families,
metric_key="family_max",
ylabel="Ξ max score (GRPO β same-size base)",
title="(b) Peak capability\npositive = GRPO unlocks higher ceiling on at least 1 scenario",
)
handles, labels = ax_mean.get_legend_handles_labels()
fig.legend(handles, labels, loc="lower center", ncol=min(4, len(pairs)), fontsize=9, bbox_to_anchor=(0.5, -0.02))
fig.suptitle("Where GRPO helps vs. hurts, per task family", fontsize=13)
fig.tight_layout(rect=[0, 0.05, 1, 0.96])
fig.savefig(out_path, dpi=160, bbox_inches="tight")
plt.close(fig)
print(f"[ok] {out_path}")
# ---------------------------------------------------------------------------
# Plot 7 β summary table as PNG
# ---------------------------------------------------------------------------
def render_summary_table(specs: dict[str, dict], out_path: Path) -> dict:
"""Render the runs_summary as a PNG (suitable for README embed) AND
return the same data as a dict so the JSON sibling can be written.
"""
families = _all_families(specs)
rows: list[dict] = []
for label, entry in specs.items():
s = entry["summary"]
row = {
"label": label,
"model": s.get("model", "?"),
"n": s.get("scenarios_total", "?"),
"avg_score": float(s.get("avg_score", 0.0)),
"format_pass_rate": float(s.get("format_pass_rate", 0.0) or 0.0),
"completion_rate": float(s.get("completion_rate", 0.0) or 0.0),
**{f"fam_{fam}": entry["family_means"].get(fam, 0.0) for fam in families},
**{f"max_{fam}": entry["family_max"].get(fam, 0.0) for fam in families},
}
rows.append(row)
summary = {"families": families, "rows": rows}
# Render text as PNG
import matplotlib.pyplot as plt
from matplotlib import patches
headers = ["Run", "n", "avg", "fmt%"] + [fam for fam in families]
body: list[list[str]] = []
for row in rows:
body.append([
row["label"],
str(row["n"]),
f"{row['avg_score']:.4f}",
f"{row['format_pass_rate'] * 100:.0f}%",
*[f"{row[f'fam_{fam}']:.3f}" for fam in families],
])
n_cols = len(headers)
n_rows = len(body) + 1
char_widths = [max(len(headers[c]), max((len(b[c]) for b in body), default=0)) for c in range(n_cols)]
total_chars = sum(char_widths)
rel_widths = [w / total_chars for w in char_widths]
fig_width = max(13, sum(char_widths) * 0.13)
fig, ax = plt.subplots(figsize=(fig_width, 0.6 * n_rows + 0.8))
ax.axis("off")
table = ax.table(
cellText=body,
colLabels=headers,
loc="center",
cellLoc="center",
colLoc="center",
colWidths=rel_widths,
)
table.auto_set_font_size(False)
table.set_fontsize(10)
table.scale(1.0, 1.5)
for c in range(n_cols):
table[(0, c)].set_facecolor("#cccccc")
table[(0, c)].set_text_props(weight="bold")
# highlight winning rows per column
for c, fam in enumerate(headers):
if c < 4:
continue
col_vals = [row[f"fam_{fam}"] for row in rows]
if not col_vals:
continue
max_v = max(col_vals)
if max_v <= 0:
continue
for r, row in enumerate(rows, start=1):
if abs(row[f"fam_{fam}"] - max_v) < 1e-9:
table[(r, c)].set_facecolor("#cdeac0") # green
table[(r, c)].set_text_props(weight="bold")
ax.set_title("ClarifyRL β per-run Γ per-family scoreboard (n=50, eval v4)\nGreen cell = best score in that family", pad=14)
fig.tight_layout()
fig.savefig(out_path, dpi=160, bbox_inches="tight")
plt.close(fig)
print(f"[ok] {out_path}")
return summary
# ---------------------------------------------------------------------------
# main
# ---------------------------------------------------------------------------
def main(argv: Iterable[str] | None = None) -> None:
p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
p.add_argument("--out-dir", default="plots", help="Directory for PNGs + runs_summary.json")
args = p.parse_args(list(argv) if argv is not None else None)
out_dir = Path(args.out_dir)
out_dir.mkdir(parents=True, exist_ok=True)
specs: dict[str, dict] = {}
for spec in RUN_SPECS:
path = _resolve_eval_path(spec)
if path is None:
print(f"[skip] {spec.label}: no eval JSON yet at {spec.eval_path}")
continue
summary, results = _load_summary_and_results(path)
specs[spec.label] = {
"summary": summary,
"family_means": _per_family_means(results),
"family_max": _per_family_max(results),
"base_label": spec.base_label,
"color": spec.color,
"eval_path": str(path),
}
if not specs:
print("[err] no eval JSONs found at all β nothing to plot")
return
print(f"\n[load] {len(specs)} eval JSON(s):")
for lbl, entry in specs.items():
print(f" - {lbl}: {entry['eval_path']}")
print()
plot_same_base_delta(specs, out_dir / "06_same_base_delta.png")
summary = render_summary_table(specs, out_dir / "07_runs_summary_table.png")
json_path = out_dir / "runs_summary.json"
json_path.write_text(json.dumps(summary, indent=2))
print(f"[ok] {json_path}")
print()
print(f"All comparison artifacts written to: {out_dir.resolve()}")
if __name__ == "__main__":
main()
|