make_board_overlay.py

#!/usr/bin/env python3
"""GoLLeM-v5 vs Glint board (74 models) — where are our models? ARC / WikiText / efficiency vs params."""
import json, matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt

D = json.load(open(r"C:/Projekty/Slayer/gollem-v5-en-staging/board_eff_vs_params.json", encoding="utf-8"))
board, ours = D["board"], D["ours"]

def xs(rows, k): return [r["params_M"] for r in rows], [r[k] for r in rows]

fig, ax = plt.subplots(1, 3, figsize=(16, 4.8), dpi=130)
panels = [("arc", "ARC-Easy (higher=better)", False),
          ("wiki", "WikiText-2 byte_ppl (lower=better)", True),
          ("eff", "Efficiency = score x size-bonus (board rank metric)", False)]

for a, (key, title, low_better) in zip(ax, panels):
    bx, by = xs(board, key)
    a.scatter(bx, by, s=42, c="#9aa0a6", alpha=.65, edgecolor="none", label="Glint board (74)")
    for r in ours:
        a.scatter([r["params_M"]], [r[key]], s=300, marker="*", c="#e03131",
                  edgecolor="k", zorder=6, label="GoLLeM-v5 (ours)" if r is ours[0] else None)
        tag = "16M" if "16M" in r["name"] else "32M"
        a.annotate(f'{tag}\n{r[key]:.1f}', (r["params_M"], r[key]), textcoords="offset points",
                   xytext=(8, 6 if not low_better else -18), fontsize=8, color="#e03131", fontweight="bold")
    a.set_xscale("log"); a.set_xlabel("parameters (M, log)"); a.set_ylabel(key)
    if key == "wiki": a.set_ylim(0, 75)  # clip a broken board outlier (~4e6) to match board view
    a.set_title(title, fontsize=10); a.grid(alpha=.3, which="both")
    a.legend(fontsize=8, loc="upper left" if not low_better else "upper right")

# efficiency panel: mark frontier + note
axe = ax[2]
best = max(board, key=lambda r: r["eff"])
axe.annotate(f'#1 {best["name"][:16]}\neff {best["eff"]:.1f}', (best["params_M"], best["eff"]),
             textcoords="offset points", xytext=(-10, -28), fontsize=7, color="#1864ab")
axe.axhline(ours[0]["eff"], color="#e03131", ls=":", lw=.8, alpha=.5)
axe.annotate("16M > 32M on eff\n(size-bonus penalty > ARC gain)\n-> 16M optimal for THIS bench",
             (0.5, 0.04), xycoords="axes fraction", fontsize=7.5, ha="center", color="#555")

fig.suptitle("GoLLeM-v5 vs Glint Tiny-ML board — our models (red stars) among 74 entries", fontsize=13, y=1.03)
fig.tight_layout()
out = "C:/Projekty/datasets/gollem-v5-models/board_overlay_v5.png"
fig.savefig(out, bbox_inches="tight")
print("saved", out, "| board n=", len(board), "| ours=", [o["name"] for o in ours])
print("eff range board:", round(min(r["eff"] for r in board),2), "-", round(max(r["eff"] for r in board),2))