File size: 5,444 Bytes
bf5f52e | 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 | """Generate Plotly HTML figures (CDN, self-contained data) for the logbook."""
import json, math, sys
import plotly.graph_objects as go
def fig_noise():
d = json.load(open("noise_audit_results.json"))
r = d["rows"]
B = [x["B"] for x in r]
fig = go.Figure()
fig.add_scatter(x=B, y=[x["N3_median"] for x in r], mode="lines+markers",
name="measured N₃ (median)")
fig.add_scatter(x=B, y=[x["pred_Eq1"] for x in r], mode="lines+markers",
name="(B/d_model)·√(d_k log m) [Eq.1, up to const]", line=dict(dash="dash"))
fig.add_scatter(x=B, y=[x["signal"] for x in r], mode="lines+markers",
name="true-edge signal (≈ d_k, block-independent)", line=dict(dash="dot"))
fig.add_hline(y=d["d_k"] / 2, line=dict(color="red", dash="dot"),
annotation_text="τ = d_k/2 (separation fails above)")
fig.update_layout(title=f"Claim 3 — superposition noise N₃(B) grows linearly in block size B "
f"(m={d['m']}, d_model={d['d_model']}, d_k={d['d_k']})",
xaxis_title="block size B (= targets served by a head)",
yaxis_title="score", xaxis_type="log", yaxis_type="log",
template="plotly_white", legend=dict(y=0.02, x=0.02))
fig.write_html("fig_claim3.html", include_plotlyjs="cdn")
# raw csv
with open("fig_claim3.csv", "w") as f:
f.write("B,signal,N3_median,N3_max,pred_Eq1\n")
for x in r:
f.write(f"{x['B']},{x['signal']:.2f},{x['N3_median']:.2f},{x['N3_max']:.2f},{x['pred_Eq1']:.2f}\n")
print("wrote fig_claim3.html / .csv")
def fig_claim4():
d = json.load(open("results_base.json"))
S = [r for r in d["summary"] if "D_K_star" in r]
x = [r["x"] for r in S]; y = [r["D_K_star"] for r in S]
slope = d.get("fit", {}).get("slope"); r2 = d.get("fit", {}).get("r2")
fig = go.Figure()
fig.add_scatter(x=x, y=y, mode="markers+text",
text=[f"({r['m']},{r['d_model']})" for r in S],
textposition="top center", name="D_K* (measured)",
marker=dict(size=10, color=["red" if r["d_model"] == 16 and r["m"] > 64 else "royalblue" for r in S]))
xs = [0, max(x) * 1.05]
if slope:
fig.add_scatter(x=xs, y=[slope * v for v in xs], mode="lines",
name=f"fit: D_K*={slope:.2f}·x (R²={r2:.3f})", line=dict(dash="dash"))
fig.update_layout(title="Claim 4 — empirical capacity law D_K* vs m·log(m)/d_model "
"(red = d_model=16, m>64 outliers)",
xaxis_title="x = m·log m / d_model", yaxis_title="D_K* (min key dim for 0.99 F1)",
template="plotly_white")
fig.write_html("fig_claim4.html", include_plotlyjs="cdn")
with open("fig_claim4.csv", "w") as f:
f.write("m,d_model,x,D_K_star,h_star,d_k_star\n")
for r in S:
f.write(f"{r['m']},{r['d_model']},{r['x']:.2f},{r['D_K_star']},{r.get('h_star')},{r.get('d_k_star')}\n")
print(f"wrote fig_claim4.html / .csv (n={len(S)})")
def fig_claim5():
d = json.load(open("results_gpt2.json"))
agg = {}
for r in d["records"]:
agg.setdefault((r["D_K"], r["h"]), []).append(r["test_acc"])
DKs = sorted({k[0] for k in agg}); hs = sorted({k[1] for k in agg})
fig = go.Figure()
for h in hs:
xs = [dk for dk in DKs if (dk, h) in agg]
ys = [sum(agg[(dk, h)]) / len(agg[(dk, h)]) for dk in xs]
fig.add_scatter(x=xs, y=ys, mode="lines+markers", name=f"h={h}")
fig.update_layout(title="Claim 5 — GPT-2 block: test accuracy vs D_K by head count "
f"(m={d['config']['m']}, d_model=768, ℓ={d['config']['ell']})",
xaxis_title="D_K (total key dim)", yaxis_title="test accuracy",
template="plotly_white")
fig.write_html("fig_claim5.html", include_plotlyjs="cdn")
with open("fig_claim5.csv", "w") as f:
f.write("D_K,h,mean_test_acc,n\n")
for (dk, h), v in sorted(agg.items()):
f.write(f"{dk},{h},{sum(v)/len(v):.4f},{len(v)}\n")
print("wrote fig_claim5.html / .csv")
def fig_claim6():
d = json.load(open("results_claim6.json"))
fig = go.Figure()
for (m, dm) in sorted({(r["m"], r["d_model"]) for r in d}):
rows = [r for r in d if r["m"] == m and r["d_model"] == dm]
rows.sort(key=lambda r: r["D_K"])
xs = [r["D_K"] for r in rows]
for tag, name in [("t16e16", "train16/eval16"), ("t32e32", "train32/eval32"), ("t16e32", "train16/eval32")]:
fig.add_scatter(x=xs, y=[r[tag] for r in rows], mode="lines+markers",
name=f"({m},{dm}) {name}")
fig.update_layout(title="Claim 6 — capacity threshold vs context length ℓ∈{16,32} and length generalization",
xaxis_title="D_K", yaxis_title="test micro-F1", template="plotly_white")
fig.write_html("fig_claim6.html", include_plotlyjs="cdn")
print("wrote fig_claim6.html")
if __name__ == "__main__":
which = sys.argv[1] if len(sys.argv) > 1 else "all"
fns = {"noise": fig_noise, "claim4": fig_claim4, "claim5": fig_claim5, "claim6": fig_claim6}
if which == "all":
for f in fns.values():
try: f()
except Exception as e: print("skip", f.__name__, e)
else:
fns[which]()
|