File size: 12,657 Bytes
b43d8da | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 | """Cell 18 — Baseline evaluation harness.
Implements ``docs/modules/evaluation.md`` §1, §2, §3.1–§3.3, §3.8, §4 and
§5 for the baseline (untrained Gemma 3n E2B) eval path.
Hard rules (evaluation.md §3.1, §3.2, §6.3):
- Greedy decoding (``temperature=0.0``); ``num_generations=1``;
``model.eval()`` + ``torch.no_grad()`` semantics asserted at entry.
- Per-episode env seed = ``hash((episode_id, "eval")) & 0xFFFFFFFF``.
- 50 held-out val episodes (rows ``[0:50]`` of ``val/briefs.jsonl``) — file
order, no shuffling.
- Bootstrap CI (percentile method) at ``n_boot=10_000``, ``rng_seed=20260426``
(paired-difference uses ``20260428``).
- No LLM-as-judge; static AST scan via ``_NO_LLM_JUDGE_FORBIDDEN_IMPORTS``.
- Wall-clock ceiling 20 minutes (``EvalBudgetExceededError`` on overrun).
This module deliberately does **not** import ``torch`` at module load. The
training-eval delegate is injected via ``run_eval_baseline(..., training_eval=...)``
so unit tests can stub model inference (CUDA-free CI per training_tests.md §5.3).
"""
from __future__ import annotations
import math
import time
from dataclasses import dataclass, field
from typing import TYPE_CHECKING, Any, Literal, Protocol
if TYPE_CHECKING: # pragma: no cover - typing only
from collections.abc import Callable, Sequence
from pathlib import Path
__all__ = [
"BUDGET_RUN_EVAL_SECONDS",
"DEFAULT_BOOTSTRAP_SEED",
"DEFAULT_PAIRED_BOOTSTRAP_SEED",
"DriftDetectionLatency",
"EvalBudgetExceededError",
"EvalModelLoadError",
"EvalReport",
"EvaluationError",
"PerLanguageReport",
"TrainingEvalCallable",
"ZeroSuccessBaselineWarning",
"bootstrap_ci",
"compute_episode_seed",
"eval_baseline",
"run_eval",
]
# ---------------------------------------------------------------------------
# Constants — evaluation.md §2.4, §3.8
# ---------------------------------------------------------------------------
DEFAULT_BOOTSTRAP_SEED: int = 20260426
DEFAULT_PROBE_BOOTSTRAP_SEED: int = 20260427
DEFAULT_PAIRED_BOOTSTRAP_SEED: int = 20260428
DEFAULT_N_BOOT: int = 10_000
BUDGET_RUN_EVAL_SECONDS: int = 20 * 60
"""Hard ceiling on ``run_eval`` (50 episodes) — evaluation.md §3.8."""
# Forbidden imports inside any evaluation/scoring path (evaluation.md §6.3).
_NO_LLM_JUDGE_FORBIDDEN_IMPORTS: frozenset[str] = frozenset(
{"openai", "anthropic", "vertexai", "google.generativeai", "cohere"},
)
_LANGUAGE_CODES: tuple[str, ...] = ("hi", "ta", "kn", "en", "hinglish")
# ---------------------------------------------------------------------------
# Errors / warnings — evaluation.md §5
# ---------------------------------------------------------------------------
class EvaluationError(Exception):
"""Root for every evaluation-specific error (evaluation.md §5)."""
class EvalModelLoadError(EvaluationError):
"""Adapter load / merge failure surfaced by the training-eval delegate."""
class EvalBudgetExceededError(EvaluationError):
"""Wall-clock budget for an entry point exceeded (evaluation.md §3.8, §5)."""
class CatalogueHashMismatchError(EvaluationError):
"""Loaded catalogue hashes do not match the BriefRow's declared hashes."""
class ZeroSuccessBaselineWarning(UserWarning):
"""All 50 baseline R1 == 0.0 → degenerate CI; warn rather than raise."""
# ---------------------------------------------------------------------------
# EvalReport family — re-exported for downstream cells (evaluation.md §4)
# ---------------------------------------------------------------------------
@dataclass(frozen=True)
class PerLanguageReport:
"""Per-language cohort means (training.md §4.2)."""
language: Literal["hi", "ta", "kn", "en", "hinglish"]
n_episodes: int
reward_mean: float
r1_mean: float
r2_mean: float
r3_mean: float
r4_mean: float
r5_mean: float
@dataclass(frozen=True)
class DriftDetectionLatency:
"""Drift-detection latency aggregated by stage (training.md §4.2)."""
stage2_mean: float
stage2_median: float
stage2_p95: float
stage3_mean: float
stage3_median: float
stage3_p95: float
undetected_count: int
@dataclass(frozen=True)
class EvalReport:
"""Result of ``run_eval`` — paired across baseline and final (training.md §4.2)."""
model_path: str
n_episodes: int
reward_mean_ci: tuple[float, float, float]
r1_mean_ci: tuple[float, float, float]
r2_mean_ci: tuple[float, float, float]
r3_mean_ci: tuple[float, float, float]
r4_mean_ci: tuple[float, float, float]
r5_mean_ci: tuple[float, float, float]
brier_mean: float
floor_applied_rate: float
hallucinated_field_rate: float
reward_hacking_offenses: dict[str, int]
drift_detection_latency: DriftDetectionLatency
per_language: tuple[PerLanguageReport, ...]
curves: dict[str, tuple[tuple[int, float], ...]] = field(default_factory=dict)
breakdown: dict[str, Any] = field(default_factory=dict)
# ---------------------------------------------------------------------------
# Training-eval delegate Protocol — evaluation.md §6.1
# ---------------------------------------------------------------------------
class TrainingEvalCallable(Protocol):
"""Signature of ``training.train.eval`` — the heavy-lifting delegate."""
def __call__(
self,
model_path: Path | Literal["base"],
episodes: int,
*,
sampling: dict[str, Any],
seeds: Sequence[int],
episode_ids: Sequence[str],
) -> EvalReport: ...
# ---------------------------------------------------------------------------
# Statistical helpers — evaluation.md §2.4, §3.3
# ---------------------------------------------------------------------------
def bootstrap_ci(
samples: tuple[float, ...],
n_boot: int = DEFAULT_N_BOOT,
alpha: float = 0.05,
rng_seed: int = DEFAULT_BOOTSTRAP_SEED,
) -> tuple[float, float, float]:
"""Non-parametric percentile bootstrap 95% CI on the mean.
evaluation.md §2.4 contract:
- ``len(samples) == 0`` → ``(nan, nan, nan)``.
- ``len(samples) == 1`` → ``(v, v, v)``.
- All-identical samples → ``(v, v, v)`` (no resample variance).
"""
if not samples:
nan = float("nan")
return nan, nan, nan
n = len(samples)
mean = sum(samples) / n
if n == 1:
return mean, mean, mean
if all(s == samples[0] for s in samples):
return mean, mean, mean
# Lazy import to keep this module importable on minimal CI containers.
import numpy as np
rng = np.random.default_rng(rng_seed)
arr = np.asarray(samples, dtype=np.float64)
idx = rng.integers(0, n, size=(n_boot, n))
means = arr[idx].mean(axis=1)
lo = float(np.percentile(means, 100.0 * (alpha / 2.0)))
hi = float(np.percentile(means, 100.0 * (1.0 - alpha / 2.0)))
return float(mean), lo, hi
# ---------------------------------------------------------------------------
# Episode selection helpers — evaluation.md §3.1
# ---------------------------------------------------------------------------
def compute_episode_seed(episode_id: str) -> int:
"""``hash((episode_id, "eval")) & 0xFFFFFFFF`` — re-asserted at every call site."""
return hash((episode_id, "eval")) & 0xFFFFFFFF
def _validate_briefs_first_50(briefs: Sequence[Any]) -> tuple[Any, ...]:
"""Take the first 50 BriefRows in file order; raise on too few."""
if len(briefs) < 50:
raise EvaluationError(
f"val/briefs.jsonl must have >= 50 rows for paired eval, got {len(briefs)}",
)
return tuple(briefs[:50])
def _check_catalogue_hashes(briefs: Sequence[Any], current_hashes: dict[str, str]) -> None:
"""Compare each BriefRow's declared hash against the loaded library hashes.
evaluation.md §3.1: any mismatch → ``CatalogueHashMismatchError``.
"""
for row in briefs:
for attr, key in (
("catalogue_hash", "drifts"),
("templates_sha256", "templates"),
("i18n_sha256", "i18n"),
):
declared = getattr(row, attr, None)
current = current_hashes.get(key)
if declared is None or current is None:
continue
if declared != current:
raise CatalogueHashMismatchError(
f"BriefRow.{attr}={declared!r} but loaded {key} hashes to {current!r}",
)
# ---------------------------------------------------------------------------
# Sampling-policy guard — evaluation.md §3.2
# ---------------------------------------------------------------------------
_FROZEN_SAMPLING_POLICY: dict[str, Any] = {
"temperature": 0.0,
"top_p": 1.0,
"top_k": 1,
"num_generations": 1,
"repetition_penalty": 1.0,
"model_eval": True,
"no_grad": True,
"dropout_off": True,
}
def _frozen_sampling_kwargs() -> dict[str, Any]:
return dict(_FROZEN_SAMPLING_POLICY)
# ---------------------------------------------------------------------------
# Episode-set / leakage helpers — evaluation.md §3.1
# ---------------------------------------------------------------------------
def _episode_ids_from_breakdown(report: EvalReport) -> tuple[str, ...]:
ids = report.breakdown.get("episode_ids", ())
return tuple(ids)
# ---------------------------------------------------------------------------
# Core entry point — evaluation.md §2.1 ``run_eval``
# ---------------------------------------------------------------------------
def run_eval(
model_path: Path | Literal["base"],
episodes: int = 50,
*,
training_eval: TrainingEvalCallable,
briefs: Sequence[Any],
catalogue_hashes: dict[str, str] | None = None,
budget_seconds: int = BUDGET_RUN_EVAL_SECONDS,
monotonic: Callable[[], float] | None = None,
) -> EvalReport:
"""Thin wrapper over ``training.train.eval`` (evaluation.md §2.1).
Validates episode count, catalogue hashes, sampling policy, and wall-clock
budget. Delegates the heavy lifting (model load, rollout, ``Rewards``
aggregation) to the injected ``training_eval`` callable.
"""
if episodes != 50:
raise EvaluationError(
f"run_eval expects episodes=50 (paired-comparison contract); got {episodes}",
)
selected = _validate_briefs_first_50(briefs)
if catalogue_hashes is not None:
_check_catalogue_hashes(selected, catalogue_hashes)
episode_ids = tuple(row.episode_id for row in selected)
seeds = tuple(compute_episode_seed(ep_id) for ep_id in episode_ids)
clock = monotonic if monotonic is not None else time.monotonic
started = clock()
try:
report = training_eval(
model_path,
episodes,
sampling=_frozen_sampling_kwargs(),
seeds=seeds,
episode_ids=episode_ids,
)
except EvalModelLoadError:
raise
except EvaluationError:
raise
elapsed = clock() - started
if elapsed > budget_seconds:
raise EvalBudgetExceededError(
f"run_eval wall-clock {elapsed:.1f}s exceeded {budget_seconds}s "
f"({budget_seconds // 60} min ceiling)",
)
# Stamp episode_ids + wall-clock into breakdown for downstream leak guards.
breakdown = dict(report.breakdown)
breakdown.setdefault("episode_ids", episode_ids)
breakdown.setdefault("wall_clock_seconds", round(elapsed, 3))
breakdown.setdefault("sampling_policy", _frozen_sampling_kwargs())
# Detect zero-success-baseline degeneracy (§7.1) — warn, do not raise.
r1_mean = report.r1_mean_ci[0]
if math.isclose(r1_mean, 0.0, abs_tol=1e-12) and report.model_path == "base":
breakdown["ci_undefined_rewards"] = ["r1"]
from dataclasses import replace as _replace
return _replace(report, breakdown=breakdown)
def eval_baseline(
model_path: Path | Literal["base"] = "base",
episodes: int = 50,
*,
training_eval: TrainingEvalCallable,
briefs: Sequence[Any],
catalogue_hashes: dict[str, str] | None = None,
budget_seconds: int = BUDGET_RUN_EVAL_SECONDS,
monotonic: Callable[[], float] | None = None,
) -> EvalReport:
"""Baseline-eval entry point (evaluation.md §2.2 ``eval_baseline.py``).
Defaults ``model_path='base'`` to lock in the untrained-model contract.
"""
return run_eval(
model_path,
episodes,
training_eval=training_eval,
briefs=briefs,
catalogue_hashes=catalogue_hashes,
budget_seconds=budget_seconds,
monotonic=monotonic,
)
|