File size: 8,937 Bytes
708f4a3 | 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 | """
Crayon Vocabulary Training Module.
Implements Algorithm 3.1 from the XERV Crayon Engineering Treatise:
- Extract substring candidates up to SIMD limit (16 bytes)
- Calculate information gain with entropy reduction
- Select top-K candidates maximizing gain-to-cost ratio
This is the production-grade implementation for building optimal vocabularies
from either user-provided corpora or the built-in default sources.
"""
import math
import logging
import string
from collections import defaultdict
from typing import List, Tuple, Dict, Iterator, Optional, Callable
# Configure module logger
logger = logging.getLogger(__name__)
# SIMD Hardware Limit [cite: 128]
MAX_TOKEN_LENGTH = 16
# Minimum frequency threshold to filter noise
DEFAULT_MIN_FREQUENCY = 2
def build_default_vocabulary(
target_size: int = 500000,
progress_callback: Optional[Callable[[str], None]] = None
) -> List[str]:
"""
Builds a 'Batteries-Included' vocabulary using Xerv-AI's curated datasets.
Sources:
- Xerv-AI/GRAD (Graduate Mathematics)
- Xerv-AI/Physics-dataset-700 (Scientific Reasoning)
- Xerv-AI/RainDrop-DTS (General Instruction)
- Tiny Shakespeare (Classical Literature)
- Built-in corpus (Baseline Coverage)
No local files are required; data is streamed directly into the entropy engine.
Args:
target_size: Maximum vocabulary size (default 500k)
progress_callback: Optional callback for progress updates
Returns:
List of token strings ordered by utility
"""
from .resources import get_default_corpus_iterator
if progress_callback:
progress_callback("Initializing default corpus stream...")
corpus_stream = get_default_corpus_iterator()
return train_vocabulary(
corpus_stream,
target_size=target_size,
progress_callback=progress_callback
)
def train_vocabulary(
corpus_iterator: Iterator[str],
target_size: int = 500000,
min_frequency: int = DEFAULT_MIN_FREQUENCY,
progress_callback: Optional[Callable[[str], None]] = None
) -> List[str]:
"""
Constructs an optimal vocabulary from a corpus using first-principles entropy analysis.
Algorithm 3.1 [cite: 127-135]:
1. Extract all substrings up to MAX_TOKEN_LENGTH (16 bytes for AVX2).
2. Calculate Information Gain: Gain(s) = Frequency(s) × Entropy(s) - Cost(s).
3. Select Top-K candidates maximizing utility score.
Args:
corpus_iterator: Iterator yielding chunks/lines of text
target_size: Maximum vocabulary size (default 500k)
min_frequency: Minimum token frequency threshold
progress_callback: Optional callback for progress updates
Returns:
List of token strings ordered for stable ID assignment
"""
if progress_callback:
progress_callback("Starting Entropy-Guided Vocabulary Construction...")
logger.info("Starting Entropy-Guided Vocabulary Construction...")
# ========================================================================
# Phase 1: Candidate Extraction & Frequency Counting [cite: 128]
# ========================================================================
candidates: Dict[str, int] = defaultdict(int)
total_chars = 0
chunk_count = 0
# Process stream chunk by chunk (Zero-Disk Accumulation)
for text_chunk in corpus_iterator:
if not text_chunk:
continue
text_len = len(text_chunk)
total_chars += text_len
chunk_count += 1
# Hot-path extraction loop - extract all valid substrings
for i in range(text_len):
# Hardware constraint: Tokens > 16 bytes degrade SIMD performance
limit = min(i + MAX_TOKEN_LENGTH, text_len)
for j in range(i + 1, limit + 1):
token = text_chunk[i:j]
# Skip tokens that exceed byte limit when encoded
if len(token.encode('utf-8')) <= MAX_TOKEN_LENGTH:
candidates[token] += 1
# Progress update every 100 chunks
if chunk_count % 100 == 0 and progress_callback:
progress_callback(f"Processed {chunk_count} chunks, {len(candidates):,} candidates...")
if progress_callback:
progress_callback(f"Extracted {len(candidates):,} unique candidates from {total_chars:,} chars")
logger.info(f"Extracted {len(candidates):,} unique candidates from {total_chars:,} chars.")
# ========================================================================
# Phase 2: Information Gain Calculation [cite: 129-134]
# ========================================================================
if progress_callback:
progress_callback("Scoring candidates by information gain...")
scored_candidates: List[Tuple[str, float]] = []
for token, freq in candidates.items():
# Filter low-frequency noise
if freq < min_frequency:
continue
# Skip control characters and empty strings
if not token or not token.isprintable():
continue
# Probability p(s)
p_s = freq / total_chars
if p_s <= 0:
continue
# Information content (entropy reduction) [cite: 131]
# H(s) = -log2(p(s))
entropy = -math.log2(p_s)
# Computational Cost Estimate [cite: 133]
# Cost is linear to byte length + constant overhead for SIMD alignment
byte_length = len(token.encode('utf-8'))
comp_cost = byte_length * 0.1 + 1.0
# Information Gain [cite: 134]
# Gain = (Entropy × Frequency) / Cost
gain = (entropy * freq) / comp_cost
scored_candidates.append((token, gain))
if progress_callback:
progress_callback(f"Scored {len(scored_candidates):,} viable candidates")
logger.info(f"Scored {len(scored_candidates):,} viable candidates")
# ========================================================================
# Phase 3: Selection with Priority Categories [cite: 1009-1012]
# ========================================================================
if progress_callback:
progress_callback("Building final vocabulary...")
# Sort by gain descending
scored_candidates.sort(key=lambda x: x[1], reverse=True)
# Build vocabulary with reserved categories
vocab_set: set = set()
# 1. Special tokens (MANDATORY) [cite: 1009]
specials = ["<PAD>", "<UNK>", "<BOS>", "<EOS>"]
for s in specials:
vocab_set.add(s)
# 2. ASCII printable characters (BASELINE) [cite: 1010]
for c in string.printable:
if c not in vocab_set and c.strip():
vocab_set.add(c)
# 3. Common single-byte sequences
for i in range(256):
try:
char = chr(i)
if char.isprintable() and char not in vocab_set:
vocab_set.add(char)
except (ValueError, UnicodeDecodeError):
pass
# 4. Fill remainder with entropy-optimized tokens
remaining_slots = target_size - len(vocab_set)
added_count = 0
for token, gain in scored_candidates:
if added_count >= remaining_slots:
break
if token not in vocab_set:
vocab_set.add(token)
added_count += 1
final_vocab = list(vocab_set)
if progress_callback:
progress_callback(f"Final vocabulary: {len(final_vocab):,} tokens")
logger.info(f"Final vocabulary: {len(final_vocab):,} tokens")
return final_vocab
def calculate_corpus_entropy(corpus_iterator: Iterator[str]) -> float:
"""
Calculate Shannon entropy of a corpus [cite: 93-96].
H(X) = -Σ p(x) log2(p(x))
Args:
corpus_iterator: Iterator yielding text chunks
Returns:
Entropy in bits per character
"""
char_counts: Dict[str, int] = defaultdict(int)
total = 0
for chunk in corpus_iterator:
for char in chunk:
char_counts[char] += 1
total += 1
if total == 0:
return 0.0
entropy = 0.0
for count in char_counts.values():
p = count / total
if p > 0:
entropy -= p * math.log2(p)
return entropy
def estimate_optimal_vocab_size(entropy: float, epsilon: float = 0.5) -> int:
"""
Calculate optimal vocabulary size from corpus entropy [cite: 94].
V_optimal ≈ 2^(H(corpus) + ε)
For English text (H ≈ 1.2 bits/char), this yields ~500k tokens.
Args:
entropy: Corpus entropy in bits per character
epsilon: Adjustment factor (default 0.5)
Returns:
Estimated optimal vocabulary size
"""
return int(2 ** (entropy + epsilon))
|