File size: 10,379 Bytes
784f572 | 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 | """CLI for evaluating fixed-prefix conditional generation artifacts."""
from __future__ import annotations
import argparse
from dataclasses import asdict
import json
import math
from pathlib import Path
from typing import Sequence
from dlb.io import (
atomic_json_write,
expected_conditional_schedule,
read_conditional_samples,
sha256_file,
validate_conditional_samples,
)
from dlb.schema import ConditionalSampleRecord
from .conditional_perplexity import (
TokenizerAssets,
conditional_texts,
compute_conditional_gen_ppl,
load_offline_tokenizer,
resolve_dataset_tokenizer_assets,
)
from .generative_perplexity import (
load_offline_gpt2_large,
resolve_gpt2_assets,
)
from .self_bleu import SelfBleuConfig, compute_self_bleu
from .unigram_entropy import mean_unigram_entropy
PRODUCTION_SAMPLE_COUNT = 2048
EVALUATION_CONTINUATION_LENGTH = 64
DATASET_CONTRACTS = {
"lm1b": {"sequence_length": 128, "vocab_size": 30_522, "padding_ids": frozenset({0})},
"owt": {"sequence_length": 1024, "vocab_size": 50_257, "padding_ids": frozenset()},
}
METRIC_NAMES = (
"conditional_gen_ppl",
"mauve_suffix",
"entropy",
"self_bleu",
"prefix_exact_match",
)
def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--samples", type=Path, required=True)
parser.add_argument("--metrics", default=",".join(METRIC_NAMES))
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--root", type=Path, default=Path(__file__).resolve().parents[1])
parser.add_argument("--dataset", choices=tuple(DATASET_CONTRACTS), required=True)
parser.add_argument("--batch-size", type=int, default=8)
parser.add_argument("--context-length", type=int, default=1024)
parser.add_argument("--device", default="cuda")
parser.add_argument("--mauve-device-id", type=int, default=0)
return parser.parse_args(argv)
def _metrics(value: str) -> tuple[str, ...]:
names = tuple(part.strip() for part in value.split(",") if part.strip())
if not names:
raise ValueError("at least one metric is required")
unknown = sorted(set(names) - set(METRIC_NAMES))
if unknown:
raise ValueError(f"unknown metrics: {', '.join(unknown)}")
if len(set(names)) != len(names):
raise ValueError("metric names must not be duplicated")
return names
def _display_path(path: Path, root: Path) -> str:
try:
return path.resolve().relative_to(root.resolve()).as_posix()
except ValueError:
return path.resolve().as_posix()
def _load_records(path: Path) -> list[ConditionalSampleRecord]:
return list(read_conditional_samples(path))
def _suffix_token_records(
records: Sequence[ConditionalSampleRecord], *, reference: bool
) -> list[dict[str, object]]:
key = "reference_token_ids" if reference else "continuation_token_ids"
rows = []
for record in records:
tokens = list(getattr(record, key))[:EVALUATION_CONTINUATION_LENGTH]
rows.append({"token_ids": tokens})
return rows
def _compute_mauve(
generated: Sequence[str],
reference: Sequence[str],
*,
device_id: int,
) -> dict[str, object]:
try:
import mauve
except ImportError as error:
raise RuntimeError("mauve is required for conditional MAUVE evaluation") from error
result = mauve.compute_mauve(
p_text=list(generated),
q_text=list(reference),
device_id=device_id,
verbose=False,
)
score = float(result.mauve)
if not math.isfinite(score) or score < 0 or score > 1:
raise ValueError("MAUVE score is outside [0, 1]")
return {
"score": score,
"sample_count": len(generated),
"comparison": "generated_suffix_vs_reference_suffix",
"slice": "source_tokens_64_128_only",
"prefix_included": False,
}
def _gpt2_tokenized_self_bleu(
records: Sequence[ConditionalSampleRecord],
suffix_texts: Sequence[str],
tokenizer: object,
) -> dict[str, object]:
groups: dict[int, list[str]] = {}
for record, suffix in zip(records, suffix_texts, strict=True):
if record.prompt_id >= 256:
continue
encoded = tokenizer( # type: ignore[operator]
suffix,
add_special_tokens=False,
return_attention_mask=False,
return_token_type_ids=False,
)
token_ids = encoded["input_ids"]
if not isinstance(token_ids, list) or not all(type(token) is int for token in token_ids):
raise ValueError("GPT-2 tokenizer returned invalid Self-BLEU tokens")
groups.setdefault(record.prompt_id, []).append(" ".join(map(str, token_ids)))
if sorted(groups) != list(range(256)) or any(len(group) != 5 for group in groups.values()):
raise ValueError("conditional Self-BLEU requires the canonical 256 prompts x 5 completions")
scores = [
compute_self_bleu(groups[prompt_id], SelfBleuConfig()).score
for prompt_id in range(256)
]
return {
"score": math.fsum(scores) / len(scores),
"prompt_count": 256,
"completions_per_prompt": 5,
"sample_count": 256 * 5,
"tokenization": "gpt2",
"ngram_order": 4,
"weights": (0.25, 0.25, 0.25, 0.25),
"smoothing": "chen_cherry_method1",
"aggregation": "mean_prompt_group_self_bleu",
"slice": "generated_source_tokens_64_128_only",
}
def evaluate(args: argparse.Namespace) -> dict[str, object]:
root = args.root.resolve()
metric_names = _metrics(args.metrics)
contract = DATASET_CONTRACTS[args.dataset]
sample_count = validate_conditional_samples(
args.samples,
expected=PRODUCTION_SAMPLE_COUNT,
schedule=expected_conditional_schedule(),
sequence_length=int(contract["sequence_length"]),
vocab_size=int(contract["vocab_size"]),
)
records = _load_records(args.samples)
if sample_count != len(records):
raise ValueError("conditional sample count changed while reading records")
result_metrics: dict[str, object] = {}
dataset_tokenizer = None
decoded = None
if any(name in metric_names for name in ("conditional_gen_ppl", "mauve_suffix", "self_bleu")):
dataset_assets = resolve_dataset_tokenizer_assets(root, args.dataset)
dataset_tokenizer = load_offline_tokenizer(dataset_assets)
decoded = conditional_texts(
records,
dataset_tokenizer,
continuation_length=EVALUATION_CONTINUATION_LENGTH,
)
if "conditional_gen_ppl" in metric_names:
assets = resolve_gpt2_assets(root)
model, gpt2_tokenizer = load_offline_gpt2_large(assets, device=args.device)
ppl = compute_conditional_gen_ppl(
records,
model,
gpt2_tokenizer,
dataset_tokenizer,
batch_size=args.batch_size,
max_length=args.context_length,
continuation_length=EVALUATION_CONTINUATION_LENGTH,
model_revision=assets.model_revision,
tokenizer_revision=assets.tokenizer_revision,
device=args.device,
)
result_metrics["conditional_generative_perplexity"] = {
**asdict(ppl),
"model_snapshot": _display_path(assets.model_path, root),
"tokenizer_snapshot": _display_path(assets.tokenizer_path, root),
"prompt_loss_excluded": True,
"scored_slice": "source_tokens_64_128_only",
"offline": True,
}
if "mauve_suffix" in metric_names:
assert decoded is not None
result_metrics["mauve_suffix"] = _compute_mauve(
[item.generated_suffix for item in decoded],
[item.reference_suffix for item in decoded],
device_id=args.mauve_device_id,
)
if "entropy" in metric_names:
generated_entropy = mean_unigram_entropy(
_suffix_token_records(records, reference=False),
contract["padding_ids"],
)
reference_entropy = mean_unigram_entropy(
_suffix_token_records(records, reference=True),
contract["padding_ids"],
)
result_metrics["sample_entropy"] = {
"generated": asdict(generated_entropy),
"reference": asdict(reference_entropy),
"generated_minus_reference": (
generated_entropy.mean_entropy - reference_entropy.mean_entropy
),
"slice": "source_tokens_64_128_only",
}
if "self_bleu" in metric_names:
assert decoded is not None
assets = resolve_gpt2_assets(root)
gpt2_tokenizer = load_offline_tokenizer(
TokenizerAssets(
tokenizer_id=assets.tokenizer_id,
tokenizer_revision=assets.tokenizer_revision,
tokenizer_path=assets.tokenizer_path,
)
)
result_metrics["self_bleu"] = _gpt2_tokenized_self_bleu(
records,
[item.generated_suffix for item in decoded],
gpt2_tokenizer,
)
if "prefix_exact_match" in metric_names:
exact = sum(1 for record in records if record.prefix_exact_match is True)
result_metrics["prefix_exact_match"] = {
"rate": exact / len(records),
"matched": exact,
"sample_count": len(records),
"required_rate": 1.0,
}
return {
"schema_version": 1,
"protocol": "c64_zs_v1",
"samples": _display_path(args.samples, root),
"samples_sha256": sha256_file(args.samples),
"sample_count": sample_count,
"production_sample_count": PRODUCTION_SAMPLE_COUNT,
"dataset": args.dataset,
"prefix_length": 64,
"evaluation_continuation_length": EVALUATION_CONTINUATION_LENGTH,
"requested_metrics": list(metric_names),
"metrics": result_metrics,
}
def main(argv: Sequence[str] | None = None) -> int:
args = parse_args(argv)
document = evaluate(args)
atomic_json_write(args.output, document)
print(f"WROTE {args.output} samples={document['sample_count']} conditional=true")
return 0
if __name__ == "__main__":
raise SystemExit(main())
|