wassname/evil_MoE
1
0
Fork 0
mirror of https://github.com/wassname/evil_MoE.git synced 2026-06-27 19:47:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

plot: multi-seed overlay for substrate emergence (thin per-seed + bold mean)

Browse Source 
Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
This commit is contained in:
wassname
2026-05-30 21:23:57 +00:00
parent 0ea751c5bc
commit e45767effb
1 changed files with 28 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/plot_substrate.py
+28 -18
View File
@@ -68,33 +68,41 @@ def _onset(steps: np.ndarray, rate: np.ndarray) -> int | None:
return int(steps[nz[0]]) if len(nz) else None
def plot(run: dict, seed: str, out: Path) -> None:
def plot(runs: list[dict], seeds: list[str], out: Path) -> None:
"""One curve per mode. Multi-seed: thin per-seed lines + bold mean; a mode counts
as 'learned' if it lifts off 0 in the MEAN. Single-seed: bold line + onset dot."""
fig, ax = plt.subplots(figsize=(6.4, 3.8))
x = run["steps"]
n_learned = 0
L = min(len(r["steps"]) for r in runs)
x = runs[0]["steps"][:L]
n_learned = ymax = 0
for k, (mode, color) in HK.items():
y = run[k]
ax.plot(x, y, color=color, lw=2.0, solid_capstyle="round")
on = _onset(x, y)
final = np.nan_to_num(y)[-1]
stacked = np.stack([np.nan_to_num(r[k][:L]) for r in runs]) # (seeds, steps)
ymean = stacked.mean(axis=0)
ymax = max(ymax, ymean.max())
if len(runs) > 1:
for ys in stacked:
ax.plot(x, ys, color=color, lw=0.7, alpha=0.30, solid_capstyle="round")
ax.plot(x, ymean, color=color, lw=2.0, solid_capstyle="round")
on = _onset(x, ymean)
if on is not None:
n_learned += 1
ax.plot(on, np.nan_to_num(y)[x == on][0], "o", color=color, ms=5, zorder=5)
# direct end-label (no legend): mode + final rate, or "never" for the holdout
tag = f"{mode} {final*100:.0f}%" + ("" if on is not None else " (never)")
ax.annotate(tag, (x[-1], final), color=color, fontsize=8, va="center",
if len(runs) == 1:
ax.plot(on, ymean[x == on][0], "o", color=color, ms=5, zorder=5)
tag = f"{mode} {ymean[-1]*100:.0f}%" + ("" if on is not None else " (never)")
ax.annotate(tag, (x[-1], ymean[-1]), color=color, fontsize=8, va="center",
xytext=(6, 0), textcoords="offset points")
seed_lbl = f"seed {seeds[0]}" if len(seeds) == 1 else f"{len(seeds)} seeds: {','.join(seeds)}"
ax.set_xlabel("GRPO step")
ax.set_ylabel("cumulative hack rate (per mode)")
ax.set_title(f"vanilla GRPO learns {n_learned} of {len(HK)} loopholes (seed {seed})")
ax.set_ylim(-0.02, max(0.5, np.nanmax([np.nanmax(run[k]) for k in HK]) * 1.1))
ax.set_title(f"vanilla GRPO learns {n_learned} of {len(HK)} loopholes ({seed_lbl})")
ax.set_ylim(-0.02, max(0.5, ymax * 1.15))
ax.set_xlim(0, x[-1] * 1.28) # headroom for end-labels
ax.spines[["top", "right"]].set_visible(False)
ax.grid(axis="y", lw=0.4, alpha=0.4)
fig.tight_layout()
out.parent.mkdir(parents=True, exist_ok=True)
fig.savefig(out, dpi=140)
logger.info(f"wrote {out} ({n_learned}/{len(HK)} learned)")
logger.info(f"wrote {out} ({n_learned}/{len(HK)} learned, {len(seeds)} seed(s))")
def main() -> None:
@@ -102,10 +110,12 @@ def main() -> None:
ap.add_argument("logs", nargs="+", type=Path)
ap.add_argument("--out", type=Path, default=Path("out/figs/substrate_emergence.png"))
args = ap.parse_args()
# single-seed for now; first log wins (seed overlay is a later extension)
path = args.logs[0]
seed = (re.search(r"seed(\d+)", path.name) or re.search(r"s(\d+)", path.name))
plot(parse_hk(path), seed.group(1) if seed else "?", args.out)
runs, seeds = [], []
for path in args.logs:
runs.append(parse_hk(path))
m = re.search(r"seed(\d+)", path.name) or re.search(r"_s(\d+)", path.name)
seeds.append(m.group(1) if m else "?")
plot(runs, seeds, args.out)
if __name__ == "__main__":