File size: 5,419 Bytes
c7753e0 | 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 | #!/usr/bin/env python3
"""Parse sbob-cmp-q-* job logs into quant_matrix.{json,md}.
Each eval_ensemble.py run prints a final table:
val set examples loss ppl
spongebob 1434 2.0123 7.48
...
Rows = 11 val sets; we collect them per config (base / ensemble / 10 per-char).
"""
import json, re, subprocess, sys, os
NS = "sbob"
SLUGS = ["spongebob", "patrick", "squidward", "mr_krabs", "plankton",
"sandy", "mrs_puff", "karen", "gary", "pearl", "_ensemble"]
# config key -> k8s job name
CONFIGS = {"base": "sbob-cmp-q-base", "ensemble": "sbob-cmp-q-ensemble"}
for s in SLUGS[:-1]:
CONFIGS[s] = "sbob-cmp-q-" + s.replace("_", "-")
ROW_RE = re.compile(r"^(_?[a-z_]+)\s+(\d+)\s+([0-9.]+)\s+([0-9.]+)\s*$")
def logs(job):
out = subprocess.run(
["kubectl", "-n", NS, "logs", f"job/{job}", "--tail=200"],
capture_output=True, text=True)
return out.stdout
def parse(text):
rows = {}
for ln in text.splitlines():
m = ROW_RE.match(ln.strip())
if m and m.group(1) in SLUGS:
rows[m.group(1)] = {"examples": int(m.group(2)),
"loss": float(m.group(3)),
"ppl": float(m.group(4))}
return rows
def main():
matrix, missing = {}, []
for cfg, job in CONFIGS.items():
rows = parse(logs(job))
if len(rows) != len(SLUGS):
missing.append((cfg, sorted(set(SLUGS) - set(rows))))
matrix[cfg] = rows
out_dir = os.path.dirname(os.path.abspath(__file__))
with open(os.path.join(out_dir, "quant_matrix.json"), "w") as f:
json.dump(matrix, f, indent=2)
cfg_order = ["base", "ensemble"] + SLUGS[:-1]
# main table: rows = val sets, cols = configs, cell = loss, bold row-best
lines = ["# Quant matrix — masked val loss (gemma-4-31B base vs LoRAs)",
"",
"Configs: `base` = no adapter; `ensemble` = all-speaker r=64 LoRA; "
"others = dedicated per-character r=32 LoRAs. "
"Each column is one 1-GPU job evaluating all 11 val sets "
"(token-weighted mean masked NLL; lower is better). 2026-06-10.",
"",
"| val set \\ config | " + " | ".join(cfg_order) + " |",
"|---|" + "---|" * len(cfg_order)]
flags = []
for slug in SLUGS:
vals = {c: matrix.get(c, {}).get(slug, {}).get("loss") for c in cfg_order}
present = {c: v for c, v in vals.items() if v is not None}
best = min(present.values()) if present else None
cells = []
for c in cfg_order:
v = vals[c]
if v is None:
cells.append("—")
continue
if v > 5 or v < 0.5:
flags.append(f"{c} on {slug}: loss {v}")
cells.append(f"**{v:.4f}**" if v == best else f"{v:.4f}")
n = matrix.get("base", {}).get(slug, {}).get("examples", "?")
lines.append(f"| {slug} (n={n}) | " + " | ".join(cells) + " |")
# summary row: mean delta vs base across the 11 val sets, per config
deltas = []
for c in cfg_order:
ds = [matrix[c][s]["loss"] - matrix["base"][s]["loss"]
for s in SLUGS if s in matrix.get(c, {}) and s in matrix.get("base", {})]
deltas.append(f"{sum(ds)/len(ds):+.4f}" if ds else "—")
lines.append("| **mean Δ vs base (11 sets)** | " + " | ".join(deltas) + " |")
# per-character summary: base vs dedicated vs ensemble on the char's own val set
lines += ["", "## Ensemble vs dedicated, on each character's own val set", "",
"| val set | base | dedicated | ensemble | ens Δ vs base | ens Δ vs dedicated |",
"|---|---|---|---|---|---|"]
for slug in SLUGS[:-1]:
b = matrix.get("base", {}).get(slug, {}).get("loss")
d = matrix.get(slug, {}).get(slug, {}).get("loss")
e = matrix.get("ensemble", {}).get(slug, {}).get("loss")
if None in (b, d, e):
lines.append(f"| {slug} | — | — | — | — | — |")
continue
lines.append(f"| {slug} | {b:.4f} | {d:.4f} | {e:.4f} | "
f"{e-b:+.4f} | {e-d:+.4f} |")
b = matrix.get("base", {}).get("_ensemble", {}).get("loss")
e = matrix.get("ensemble", {}).get("_ensemble", {}).get("loss")
if b is not None and e is not None:
lines.append(f"| _ensemble (own set) | {b:.4f} | — | {e:.4f} | {e-b:+.4f} | — |")
# sanity checks
lines += ["", "## Sanity checks", ""]
for slug, known in [("squidward", 1.935), ("spongebob", 1.882)]:
v = matrix.get("ensemble", {}).get(slug, {}).get("loss")
if v is None:
lines.append(f"- ensemble on {slug}: MISSING (expected ~{known})")
else:
ok = abs(v - known) < 0.05
lines.append(f"- ensemble on {slug}: {v:.4f} vs known {known} -> "
+ ("OK" if ok else "MISMATCH"))
lines.append("- degenerate cells (loss >5 or <0.5): "
+ ("; ".join(flags) if flags else "none"))
if missing:
lines.append("- incomplete configs: " + "; ".join(
f"{c} missing {m}" for c, m in missing))
with open(os.path.join(out_dir, "quant_matrix.md"), "w") as f:
f.write("\n".join(lines) + "\n")
print(json.dumps({"missing": missing, "flags": flags}, indent=2))
if __name__ == "__main__":
main()
|