File size: 8,152 Bytes
68a4c53 | 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 | """PUA candidate selection.
Two strategies are provided:
* `select_by_savings` — the production path. Ranks candidates by
`frequency * (baseline_token_cost - 1)` so PUA slots go to tokens that
actually save tokens vs the baseline (cl100k). Tokens whose baseline
cost is 1 (single byte/char) score 0 — byte fallback already handles
them optimally.
* `select_by_coverage` — the legacy v0 strategy. Ranks by raw frequency.
Kept as a deprecated shim for callers that haven't migrated; emits a
`DeprecationWarning` on use.
The savings strategy also applies a tier-aware penalty: tokens beyond
`PUA_BMP_SIZE` would be assigned a 4-byte supplementary-plane PUA char
(vs 3 bytes for BMP). For tokens whose token-savings == 1, the byte cost
of substitution can erase the saving, so by default we cap the budget at
`PUA_BMP_SIZE` unless the caller opts into supplementary planes.
"""
from __future__ import annotations
import warnings
from collections import Counter
import regex as _re
from .baseline import BaselineTokenizer
from .pua import PUA_BMP_SIZE
DEFAULT_MAX_LEN = 50
# Heuristic regex for hash / UUID / base64-blob shapes — these waste PUA slots.
# Keep conservative: false positives just demote a token from PUA to byte
# fallback, which is fine.
_HEX_HASH_RE = _re.compile(r"^[0-9a-f]{16,}$", _re.IGNORECASE)
_UUID_RE = _re.compile(
r"^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$", _re.IGNORECASE
)
_BASE64_BLOB_RE = _re.compile(r"^[A-Za-z0-9+/=_\-]{24,}$")
# A "blob" must be mostly base64 alphabet with no recognizable English-word
# substrings. We approximate: if the token is long, has high alphabet entropy,
# and at least 1/3 digits or 1/3 mixed-case alpha runs, treat it as a blob.
def _looks_like_secret_or_hash(tok: str) -> bool:
"""Heuristic: True if `tok` looks like a hash, UUID, or base64 blob."""
if _UUID_RE.match(tok):
return True
if _HEX_HASH_RE.match(tok):
return True
if len(tok) >= 24 and _BASE64_BLOB_RE.match(tok):
# Has the shape — confirm it lacks lowercase-word substructure.
# A real identifier has long-ish lowercase runs; a base64 blob is
# interleaved upper/lower/digit. Require at least 3 short runs of
# contiguous lowercase letters before we accept it as "wordy".
lower_runs = _re.findall(r"[a-z]{3,}", tok)
if len(lower_runs) < 2:
return True
return False
def is_good_token(tok: str, max_len: int = DEFAULT_MAX_LEN) -> bool:
"""Filter for PUA-eligible tokens.
Rejects:
- Empty / whitespace-only / over-length tokens.
- Tokens that encode to a single UTF-8 byte (byte fallback handles them
optimally, so a PUA assignment is at best a wash).
- Hashes, UUIDs, long base64 blobs (waste PUA slots; near-zero
cross-document reuse).
"""
if not tok:
return False
if tok.isspace():
return False
if len(tok) > max_len:
return False
if len(tok.encode("utf-8")) <= 1:
return False
return not _looks_like_secret_or_hash(tok)
def compute_candidate_score(
token: str,
frequency: int,
baseline: BaselineTokenizer,
) -> float:
"""Net token savings per occurrence times frequency.
`savings = max(0, baseline.count_tokens(token) - 1)`.
A token with `baseline_count == 1` scores 0 — substituting it costs a
PUA slot but doesn't shorten the token sequence. A token with
`baseline_count == 5` and frequency 1000 scores 4000.
Note: this scoring is intentionally tier-agnostic. The tier-aware
penalty (BMP vs supplementary plane) is applied in `select_by_savings`
after ranking, not here, because the tier of any given candidate
depends on how many candidates outrank it.
"""
baseline_cost = baseline.count_tokens(token)
savings = max(0, baseline_cost - 1)
return frequency * savings
def select_by_savings(
freq: Counter[str],
baseline: BaselineTokenizer,
*,
vocab_budget: int,
max_len: int = DEFAULT_MAX_LEN,
min_score: float = 1.0,
allow_supplementary_pua: bool = False,
) -> list[str]:
"""Select PUA candidates by net token savings.
Parameters
----------
freq
Frequency counter from `count_frequencies`.
baseline
Baseline tokenizer used to score `(baseline_cost - 1)` per token.
vocab_budget
Hard upper bound on the number of selected tokens. Capped at
`PUA_BMP_SIZE` unless `allow_supplementary_pua` is True.
max_len
Reject tokens longer than this.
min_score
Reject tokens whose score is below this threshold. Default 1.0
(i.e., at least one token saved across the entire corpus).
allow_supplementary_pua
If True, the budget is uncapped (limited only by `PUA_TOTAL`).
Off by default because supplementary-plane PUA chars are 4 bytes
each and net savings can go negative for `baseline_count==2`
candidates.
Returns
-------
Tokens in deterministic order: descending score, ascending lex for ties.
"""
if vocab_budget <= 0:
return []
effective_budget = vocab_budget
if not allow_supplementary_pua:
effective_budget = min(effective_budget, PUA_BMP_SIZE)
scored: list[tuple[float, str]] = []
for token, count in freq.items():
if not is_good_token(token, max_len=max_len):
continue
score = compute_candidate_score(token, count, baseline)
if score < min_score:
continue
scored.append((score, token))
# Sort by (-score, token) for deterministic, savings-priority order.
scored.sort(key=lambda st: (-st[0], st[1]))
return [tok for _, tok in scored[:effective_budget]]
def select_by_coverage(
freq: Counter[str],
coverage_target: float = 0.90,
max_len: int = DEFAULT_MAX_LEN,
max_tokens: int | None = None,
) -> list[str]:
"""Frequency-based selection (deprecated, retained for backward compat).
Stops at the first of: cumulative coverage ≥ `coverage_target`,
`max_tokens` selected, or input exhausted. New code should use
`select_by_savings`.
"""
if not 0.0 < coverage_target <= 1.0:
raise ValueError(f"coverage_target must be in (0,1], got {coverage_target}")
total = sum(freq.values())
if total == 0:
return []
sorted_items = sorted(freq.items(), key=lambda kv: (-kv[1], kv[0]))
selected: list[str] = []
cumulative = 0
threshold = coverage_target * total
for token, count in sorted_items:
if not is_good_token(token, max_len=max_len):
continue
selected.append(token)
cumulative += count
if max_tokens is not None and len(selected) >= max_tokens:
break
if cumulative >= threshold:
break
return selected
def coverage_of(freq: Counter[str], tokens: list[str]) -> float:
"""Fraction of total frequency covered by `tokens`. Used for diagnostics."""
total = sum(freq.values())
if total == 0:
return 0.0
covered = sum(freq.get(t, 0) for t in tokens)
return covered / total
def select_by_coverage_deprecated(
freq: Counter[str],
coverage_target: float = 0.90,
max_len: int = DEFAULT_MAX_LEN,
max_tokens: int | None = None,
) -> list[str]:
"""Shim that emits a DeprecationWarning then delegates to `select_by_coverage`."""
warnings.warn(
"select_by_coverage is deprecated; use select_by_savings with a "
"BaselineTokenizer for production builds.",
DeprecationWarning,
stacklevel=2,
)
return select_by_coverage(freq, coverage_target, max_len, max_tokens)
__all__ = [
"DEFAULT_MAX_LEN",
"compute_candidate_score",
"coverage_of",
"is_good_token",
"select_by_coverage",
"select_by_coverage_deprecated",
"select_by_savings",
]
|