File size: 10,292 Bytes
22decf5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 | """Cross-run statistics: what, if anything, separated a good run from a bad one.
Accuracy is not comparable across benchmarks -- 0.95 on BFCL and 0.17 on AIME
are both ordinary -- so every comparison here works on `z`, the within-task
z-score of the final accuracy. A feature that "helps" is one that moves a run
up relative to the other runs on *its own* benchmark.
With 48 runs over 7 benchmarks these are descriptive, not inferential: n per
task is 4-8, so the right reading of any correlation is "worth a look", not
"established". Effect sizes are reported alongside n for that reason.
"""
from __future__ import annotations
import math
from collections import defaultdict
from typing import Any, Iterable, Sequence
from .features import RunFeatures
NUMERIC_FEATURES = [
"wall_h",
"turns",
"tool_calls",
"output_tokens",
"cost_usd",
"thinking_blocks",
"thinking_chars",
"narration_chars",
"n_eval_calls",
"n_scored_evals",
"n_distinct_model_versions",
"first_eval_h",
"tail_idle_h",
"last_improvement_h",
"time_awareness_calls",
"tool_errors",
"bg_tasks",
"max_parallel_bg",
"bg_busy_s",
"machine_busy_frac",
"dead_air_h",
"compactions",
"api_retries",
"mean_eval_limit",
"final_message_chars",
]
def _rank(xs: Sequence[float]) -> list[float]:
order = sorted(range(len(xs)), key=lambda i: xs[i])
ranks = [0.0] * len(xs)
i = 0
while i < len(order):
j = i
while j + 1 < len(order) and xs[order[j + 1]] == xs[order[i]]:
j += 1
avg = (i + j) / 2 + 1
for k in range(i, j + 1):
ranks[order[k]] = avg
i = j + 1
return ranks
def spearman(xs: Sequence[float], ys: Sequence[float]) -> float | None:
if len(xs) < 4:
return None
rx, ry = _rank(xs), _rank(ys)
mx, my = sum(rx) / len(rx), sum(ry) / len(ry)
num = sum((a - mx) * (b - my) for a, b in zip(rx, ry))
den = math.sqrt(sum((a - mx) ** 2 for a in rx) * sum((b - my) ** 2 for b in ry))
return num / den if den else None
def within_task_z(feats: Iterable[RunFeatures]) -> dict[str, float]:
"""Final accuracy, z-scored inside each benchmark.
Tasks with a single run get z = 0: there is nothing to be better than.
"""
by_task: defaultdict[str, list[RunFeatures]] = defaultdict(list)
for f in feats:
if f.accuracy is not None:
by_task[f.task].append(f)
z: dict[str, float] = {}
for task, group in by_task.items():
accs = [f.accuracy for f in group] # type: ignore[misc]
m = sum(accs) / len(accs)
sd = math.sqrt(sum((a - m) ** 2 for a in accs) / len(accs)) if len(accs) > 1 else 0.0
for f in group:
z[f.run_id] = ((f.accuracy - m) / sd) if sd else 0.0 # type: ignore[operator]
return z
def feature_correlations(feats: list[RunFeatures]) -> list[dict[str, Any]]:
z = within_task_z(feats)
rows = []
for name in NUMERIC_FEATURES:
pairs = [
(getattr(f, name), z[f.run_id])
for f in feats
if f.run_id in z and getattr(f, name, None) is not None
]
if len(pairs) < 8:
continue
rho = spearman([p[0] for p in pairs], [p[1] for p in pairs])
if rho is None:
continue
rows.append({"feature": name, "rho_vs_within_task_z": round(rho, 3), "n": len(pairs)})
return sorted(rows, key=lambda r: -abs(r["rho_vs_within_task_z"]))
def technique_effects(feats: list[RunFeatures], min_each: int = 4) -> list[dict[str, Any]]:
"""Mean within-task z for runs that used a technique vs runs that did not."""
z = within_task_z(feats)
names: set[str] = set()
for f in feats:
names.update(f.techniques)
rows = []
for name in sorted(names):
with_, without = [], []
for f in feats:
if f.run_id not in z:
continue
(with_ if name in f.techniques else without).append(z[f.run_id])
if len(with_) < min_each or len(without) < min_each:
continue
rows.append(
{
"technique": name,
"n_with": len(with_),
"n_without": len(without),
"mean_z_with": round(sum(with_) / len(with_), 3),
"mean_z_without": round(sum(without) / len(without), 3),
"delta_z": round(sum(with_) / len(with_) - sum(without) / len(without), 3),
}
)
return sorted(rows, key=lambda r: -r["delta_z"])
def per_task_table(feats: list[RunFeatures]) -> list[dict[str, Any]]:
by: defaultdict[str, list[RunFeatures]] = defaultdict(list)
for f in feats:
by[f.task].append(f)
def avg(xs):
xs = [x for x in xs if x is not None]
return round(sum(xs) / len(xs), 3) if xs else None
rows = []
for task, g in sorted(by.items()):
rows.append(
{
"task": task,
"n_runs": len(g),
"mean_accuracy": avg([f.accuracy for f in g]),
"mean_turns": avg([f.turns for f in g]),
"mean_eval_calls": avg([f.n_eval_calls for f in g]),
"mean_model_versions": avg([f.n_distinct_model_versions for f in g]),
"mean_train_hours": avg([f.action_time_s.get("train", 0) / 3600 for f in g]),
"mean_eval_hours": avg([f.action_time_s.get("eval_official", 0) / 3600 for f in g]),
"mean_wait_hours": avg([f.action_time_s.get("wait_poll", 0) / 3600 for f in g]),
"mean_dead_air_h": avg([f.dead_air_h for f in g]),
"mean_busy_frac": avg([f.machine_busy_frac for f in g]),
"mean_tail_idle_h": avg([f.tail_idle_h for f in g]),
"mean_cost_usd": avg([f.cost_usd for f in g]),
"mean_oom_hits": avg([f.error_signals.get("oom", 0) for f in g]),
}
)
return rows
def rerun_variance(feats: list[RunFeatures]) -> list[dict[str, Any]]:
"""Spread between repeated runs of the same (task, base model) cell.
This is the noise floor. Any claimed improvement smaller than it is not
an improvement, and it bounds what a single-seed sweep can conclude.
"""
by: defaultdict[tuple[str, str], list[RunFeatures]] = defaultdict(list)
for f in feats:
if f.accuracy is not None:
by[(f.task, f.base_model)].append(f)
rows = []
for (task, model), g in sorted(by.items()):
if len(g) < 2:
continue
accs = sorted(f.accuracy for f in g) # type: ignore[misc]
rows.append(
{
"task": task,
"base_model": model,
"n": len(g),
"min": round(accs[0], 4),
"max": round(accs[-1], 4),
"spread": round(accs[-1] - accs[0], 4),
"rel_spread": round((accs[-1] - accs[0]) / accs[-1], 3) if accs[-1] else None,
}
)
return sorted(rows, key=lambda r: -r["spread"])
def final_verification(feats: list[RunFeatures]) -> dict[str, Any]:
"""Did the run end by grading its own submission the way the harness will?
The deliverable is `final_model`, and the harness scores it with
`evaluate.py` at full sample count. A run that never does that itself is
shipping something it has not checked -- a cheap habit with a large
downside, so it is worth knowing how often it happens.
"""
verified, partial, none = [], [], []
for f in feats:
late = [
e
for e in f.evals
if e.model_path
and "final_model" in e.model_path
and f.wall_h
and e.t_rel / 3600 > 0.6 * f.wall_h
]
if any(e.limit is None or e.limit >= 150 for e in late):
verified.append(f.run_id)
elif late:
partial.append(f.run_id)
else:
none.append(f.run_id)
n = len(feats)
return {
"n": n,
"verified_at_full_scale": len(verified),
"verified_but_subsampled": len(partial),
"never_graded_own_submission": len(none),
"pct_verified": round(100 * len(verified) / n, 1) if n else None,
"never_graded_examples": none[:8],
}
def dev_to_harness_drop(feats: list[RunFeatures]) -> dict[str, Any]:
"""Best score the agent measured itself vs the score the harness gave it.
A systematic positive drop is selection on the development signal: the
agent picked whichever checkpoint happened to look best, and a fresh
grading pass gives some of that back.
"""
rows = []
for f in feats:
if f.best_observed is None or f.accuracy is None:
continue
rows.append(
{
"run_id": f.run_id,
"task": f.task,
"best_self_measured": round(f.best_observed, 4),
"harness": round(f.accuracy, 4),
"drop": round(f.best_observed - f.accuracy, 4),
}
)
drops = [r["drop"] for r in rows]
return {
"n": len(rows),
"mean_drop": round(sum(drops) / len(drops), 4) if drops else None,
"median_drop": round(sorted(drops)[len(drops) // 2], 4) if drops else None,
"n_overstated": sum(1 for d in drops if d > 0.02),
"n_understated": sum(1 for d in drops if d < -0.02),
"worst": sorted(rows, key=lambda r: -r["drop"])[:6],
}
def oom_recovery(feats: list[RunFeatures]) -> dict[str, Any]:
"""How often an out-of-memory failure appears, and whether it ended the run.
OOM is the signature failure of this task: one H100, a 4B model, and an
agent that has to guess batch size and sequence length. Recovering from
it without human help is a capability in its own right.
"""
hit = [f for f in feats if f.error_signals.get("oom")]
return {
"runs_with_oom": len(hit),
"of": len(feats),
"pct": round(100 * len(hit) / len(feats), 1) if feats else None,
"mean_oom_events_when_hit": round(sum(f.error_signals["oom"] for f in hit) / len(hit), 2) if hit else None,
"all_still_produced_a_model": all(f.accuracy is not None for f in hit),
}
|