File size: 4,072 Bytes
d775561 | 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 | """
Stage 05 (v8b): build the AIME25 4-alpha results table.
Reads the stage-04 summary JSON and emits a CSV + Markdown table with,
per alpha: accuracy, mean thinking tokens (+ reduction vs alpha=1.0),
mean chars (+ reduction), mean reflection markers (+ reduction), and
collapse rate. Reductions are relative to the alpha=1.0 baseline.
Outputs into results/:
aime25_seed{seed}_4alpha_table.csv
aime25_seed{seed}_4alpha_table.md
"""
import argparse, csv, math, os, sys
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from configs import get_config
from configs.paths import dim_paths, ensure_dirs
from src.utils import read_json
FIELDS = [
"alpha", "n", "accuracy_%", "correct", "mean_think_tokens",
"token_reduction_%", "mean_chars", "char_reduction_%",
"mean_reflection_markers", "reflection_reduction_%",
"no_boxed", "collapse_rate_%",
]
def fmt(x, nd=2):
if x is None:
return ""
try:
if math.isnan(float(x)):
return ""
except Exception:
return str(x)
return f"{float(x):.{nd}f}"
def rows_from_summary(summary):
base = summary.get("1.00")
rows = []
for a in sorted([float(k) for k in summary.keys()], reverse=True):
k = f"{a:.2f}"
s = summary[k]
def red(field):
if not base or float(base[field]) == 0:
return None
return 1 - float(s[field]) / float(base[field])
rows.append({
"alpha": k,
"n": s.get("n", ""),
"accuracy_%": fmt(100.0 * float(s.get("accuracy", 0)), 1),
"correct": s.get("n_correct", ""),
"mean_think_tokens": fmt(float(s.get("mean_think_tokens", 0)), 1),
"token_reduction_%": fmt(100.0 * red("mean_think_tokens"), 1)
if red("mean_think_tokens") is not None else "",
"mean_chars": fmt(float(s.get("mean_chars", 0)), 1),
"char_reduction_%": fmt(100.0 * red("mean_chars"), 1)
if red("mean_chars") is not None else "",
"mean_reflection_markers": fmt(float(s.get("mean_mon", 0)), 2),
"reflection_reduction_%": fmt(100.0 * red("mean_mon"), 1)
if red("mean_mon") is not None else "",
"no_boxed": s.get("n_no_boxed", ""),
"collapse_rate_%": fmt(100.0 * float(s.get("collapse_rate", 0)), 1),
})
return rows
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--dimension", default="monitoring")
ap.add_argument("--seed", type=int, default=0)
args = ap.parse_args()
ensure_dirs(args.dimension)
cfg = get_config(args.dimension)
p = dim_paths(args.dimension)
sum_path = os.path.join(p.RESULTS_DIR, f"aime25_seed{args.seed}_4alpha_summary.json")
if not os.path.exists(sum_path):
print(f"[05] missing {sum_path} — run stage 04 first."); sys.exit(1)
data = read_json(sum_path)
rows = rows_from_summary(data["summary"])
csv_path = os.path.join(p.RESULTS_DIR, f"aime25_seed{args.seed}_4alpha_table.csv")
md_path = os.path.join(p.RESULTS_DIR, f"aime25_seed{args.seed}_4alpha_table.md")
with open(csv_path, "w", newline="", encoding="utf-8") as f:
w = csv.DictWriter(f, fieldnames=FIELDS)
w.writeheader()
for r in rows:
w.writerow({k: r.get(k, "") for k in FIELDS})
with open(md_path, "w", encoding="utf-8") as f:
f.write("| " + " | ".join(FIELDS) + " |\n")
f.write("| " + " | ".join(["---"] * len(FIELDS)) + " |\n")
for r in rows:
f.write("| " + " | ".join(str(r.get(k, "")) for k in FIELDS) + " |\n")
print(f"[05] selected_layers: {data.get('selected_layers')}")
print(f"[05] wrote:\n {csv_path}\n {md_path}\n")
print("| " + " | ".join(FIELDS) + " |")
print("| " + " | ".join(["---"] * len(FIELDS)) + " |")
for r in rows:
print("| " + " | ".join(str(r.get(k, "")) for k in FIELDS) + " |")
if __name__ == "__main__":
main()
|