COLM / tokenizer_train.py
edeneldith's picture
Add lossless native SafeTensors release and correct checkpoint configuration
3a1536e verified
Raw History Blame Contribute Delete
18.9 kB
"""
tokenizer_train.py β€” Word + Character hybrid tokenizer
======================================================
Each word token = one whole word. Everything else = single characters.
No subwords. No merges. No BPE. No external dependencies (for base).
Encode: greedy longest-match at word boundaries, character fallback.
Decode: concatenate tokens directly (spaces are their own token).
Corpus-aware mode (--corpus / --hf-dataset):
Scans data for the most useful words and adds them to the vocab
in both lowercase and Sentence case forms.
Usage:
# Build base tokenizer (no corpus)
python tokenizer_train.py --output colm_tokenizer.json
# Build with corpus from plain text file(s)
python tokenizer_train.py --corpus data.txt --output colm_tokenizer.json
python tokenizer_train.py --corpus "*.txt" --output colm_tokenizer.json
# Build with corpus from HuggingFace dataset
python tokenizer_train.py --hf-dataset wikitext --hf-config wikitext-103-raw-v1 \\
--hf-split train --hf-field text --output colm_tokenizer.json
# Use in training script
from tokenizer_train import COLMTokenizer
tok = COLMTokenizer.load("colm_tokenizer.json")
ids = tok.encode("The Machine God")
text = tok.decode(ids)
"""
import json
import re
import argparse
import glob
from collections import Counter
from pathlib import Path
# =============================================================================
# TOKEN DEFINITIONS
# =============================================================================
SPECIAL_TOKENS = ["<pad>", "<bos>", "<eos>", "<unk>"]
# Whitespace (encoded as visible tokens so decode reconstructs perfectly)
WHITESPACE_TOKENS = [" ", "\n", "\t"]
# Top 100 English words β€” each is one token
WORD_TOKENS = [
"the", "be", "to", "of", "and", "in", "that", "have", "it", "for",
"not", "on", "with", "he", "as", "you", "do", "at", "this", "but",
"his", "by", "from", "they", "we", "say", "her", "she", "or", "an",
"will", "my", "one", "all", "would", "there", "their", "what", "so", "up",
"out", "if", "about", "who", "get", "which", "go", "me", "when", "make",
"can", "like", "time", "no", "just", "him", "know", "take", "people", "into",
"year", "your", "good", "some", "could", "them", "see", "other", "than", "then",
"now", "look", "only", "come", "it's", "over", "think", "also", "back", "after",
"use", "two", "how", "our", "work", "first", "well", "way", "even", "new",
"want", "because", "any", "these", "give", "day", "most", "us",
]
# Individual letters
UPPER = list("ABCDEFGHIJKLMNOPQRSTUVWXYZ")
LOWER = list("abcdefghijklmnopqrstuvwxyz")
# Digits
DIGITS = list("0123456789")
# Symbols (keyboard + corpus unicode)
SYMBOLS = list(set([
"+", "_", "=", ")", "(", "*", "&", "^", "%", "$",
"Β£", '"', "!", "`", "Β¬", "\\", "|", "<", ">", "/",
"?", ":", "@", "~", "#", "'", ";", ",", ".",
# Extended unicode (add corpus-specific symbols here)
"\u2013", # – en-dash
"\u2014", # β€” em-dash
"\u2018", # ' left single curly quote
"\u2019", # ' right single curly quote
"-",
]))
# =============================================================================
# CORPUS SCANNING
# =============================================================================
# Matches whole words including contractions (don't, it's, they've, etc.)
# Apostrophe variants: straight ' and curly '
WORD_RE = re.compile(r"[a-zA-Z]+(?:['\u2019][a-zA-Z]+)*")
# Minimum frequency to consider a "long" word (filters OCR junk / typos)
MIN_FREQ_FOR_LENGTH = 3
def iter_corpus_lines(corpus_path=None, hf_dataset=None, hf_config=None,
hf_split="train", hf_field="text"):
"""Yield text lines from either local file(s) or a HuggingFace dataset.
Args:
corpus_path: Glob pattern or single path for local .txt files
hf_dataset: HuggingFace dataset name (e.g. "wikitext")
hf_config: HuggingFace config name (e.g. "wikitext-103-raw-v1")
hf_split: Dataset split to use
hf_field: Column name containing text
"""
if corpus_path is not None:
# Local text file(s) β€” supports glob patterns
paths = sorted(glob.glob(corpus_path))
if not paths:
raise FileNotFoundError(f"No files matched: {corpus_path}")
print(f"[corpus] Reading {len(paths)} file(s) from: {corpus_path}")
for p in paths:
with open(p, "r", encoding="utf-8", errors="replace") as f:
for line in f:
line = line.rstrip("\n\r")
if line:
yield line
elif hf_dataset is not None:
try:
from datasets import load_dataset
except ImportError:
raise ImportError(
"HuggingFace datasets not installed. Run: "
"pip install datasets --break-system-packages"
)
print(f"[corpus] Loading HuggingFace dataset: {hf_dataset}"
f" (config={hf_config}, split={hf_split}, field={hf_field})")
ds = load_dataset(hf_dataset, hf_config, split=hf_split,
trust_remote_code=True)
for row in ds:
text = row.get(hf_field, "")
if text:
yield text
else:
return # No corpus β€” base tokenizer only
def count_corpus_words(lines_iter):
"""Count all words in the corpus. Returns a Counter of lowercase words.
Contractions are preserved as-is (don't, it's, etc.) β€” lowercased.
"""
counts = Counter()
n_lines = 0
for line in lines_iter:
words = WORD_RE.findall(line)
# Normalise: lowercase, and normalise curly apostrophes to straight
normalised = [w.lower().replace("\u2019", "'") for w in words]
counts.update(normalised)
n_lines += 1
if n_lines % 500_000 == 0:
print(f" ...scanned {n_lines:,} lines, {sum(counts.values()):,} tokens")
print(f"[corpus] Done: {n_lines:,} lines, {len(counts):,} unique words, "
f"{sum(counts.values()):,} total tokens")
return counts
def select_corpus_words(counts, existing_words, n_freq=50, n_long=50):
"""Select corpus words for the tokenizer.
Strategy:
1. Top 100 by frequency β†’ take top 50 (excluding existing vocab)
2. Top 100 by length (min freq threshold) β†’ take top 50 by length
3. Deduplicate across both lists
4. Return combined list
Args:
counts: Counter of lowercase words
existing_words: Set of words already in the tokenizer (lowercase)
n_freq: How many frequency-sorted words to keep
n_long: How many length-sorted words to keep
Returns:
(freq_words, long_words, combined) β€” all lowercase, deduplicated
"""
# Single-char words are already covered by the character tokens
eligible = {w: c for w, c in counts.items()
if len(w) > 1 and w not in existing_words}
# --- Pool 1: Top by frequency ---
freq_sorted = sorted(eligible.items(), key=lambda x: x[1], reverse=True)
freq_pool = [w for w, _ in freq_sorted[:100]]
freq_selected = freq_pool[:n_freq]
# --- Pool 2: Top by length (must meet minimum frequency) ---
long_eligible = {w: c for w, c in eligible.items()
if c >= MIN_FREQ_FOR_LENGTH}
long_sorted = sorted(long_eligible.items(), key=lambda x: len(x[0]), reverse=True)
long_pool = [w for w, _ in long_sorted[:100]]
long_selected = long_pool[:n_long]
# --- Combine and deduplicate (preserve selection order) ---
seen = set()
combined = []
# Frequency words first (they're higher-value for compression)
for w in freq_selected:
if w not in seen:
combined.append(w)
seen.add(w)
# Then length words
for w in long_selected:
if w not in seen:
combined.append(w)
seen.add(w)
print(f"[vocab] Selected {len(freq_selected)} by frequency, "
f"{len(long_selected)} by length")
print(f"[vocab] Combined (deduplicated): {len(combined)} corpus words")
# Show what we picked
if freq_selected:
print(f"[vocab] Top 5 by freq: {freq_selected[:5]}")
if long_selected:
print(f"[vocab] Top 5 by length: {long_selected[:5]} "
f"(lengths: {[len(w) for w in long_selected[:5]]})")
return freq_selected, long_selected, combined
def make_sentence_case(word):
"""Capitalise first letter only β€” as it would appear at start of sentence.
Handles contractions: don't β†’ Don't, it's β†’ It's
"""
if not word:
return word
return word[0].upper() + word[1:]
# =============================================================================
# TOKENIZER CLASS
# =============================================================================
class COLMTokenizer:
def __init__(self, token_list):
"""Build tokenizer from ordered token list."""
self.token_to_id = {token: i for i, token in enumerate(token_list)}
self.id_to_token = {i: token for i, token in enumerate(token_list)}
self.vocab_size = len(token_list)
# Separate word tokens for boundary-aware matching
self.word_tokens = set()
for token in token_list:
if len(token) > 1 and token not in SPECIAL_TOKENS:
self.word_tokens.add(token)
# Precompute max word length for bounded search
self.max_word_len = max((len(w) for w in self.word_tokens), default=0)
# Special token IDs
self.pad_id = self.token_to_id["<pad>"]
self.bos_id = self.token_to_id["<bos>"]
self.eos_id = self.token_to_id["<eos>"]
self.unk_id = self.token_to_id["<unk>"]
def encode(self, text, add_bos=False, add_eos=False):
"""Encode text to token IDs.
Greedy longest-match for word tokens at word boundaries.
Single character fallback for everything else."""
ids = []
if add_bos:
ids.append(self.bos_id)
i = 0
n = len(text)
while i < n:
# Try longest word match first (only at word boundaries)
matched = False
before_ok = (i == 0) or not text[i - 1].isalpha()
if before_ok:
hi = min(self.max_word_len, n - i)
for length in range(hi, 1, -1):
candidate = text[i:i + length]
if candidate in self.word_tokens:
after_pos = i + length
after_ok = (after_pos >= n) or not text[after_pos].isalpha()
if after_ok:
ids.append(self.token_to_id[candidate])
i += length
matched = True
break
if not matched:
ch = text[i]
ids.append(self.token_to_id.get(ch, self.unk_id))
i += 1
if add_eos:
ids.append(self.eos_id)
return ids
def decode(self, ids):
"""Decode token IDs back to text. Just concatenate."""
tokens = []
for token_id in ids:
if token_id in (self.pad_id, self.bos_id, self.eos_id):
continue
tokens.append(self.id_to_token.get(token_id, "<unk>"))
return "".join(tokens)
def save(self, path):
"""Save tokenizer as JSON."""
data = {
"version": "colm_v1",
"vocab_size": self.vocab_size,
"tokens": [self.id_to_token[i] for i in range(self.vocab_size)],
}
with open(path, "w", encoding="utf-8") as f:
json.dump(data, f, ensure_ascii=False, indent=2)
print(f"Saved tokenizer to {path}")
print(f"Vocab size: {self.vocab_size}")
@classmethod
def load(cls, path):
"""Load tokenizer from JSON."""
with open(path, "r", encoding="utf-8") as f:
data = json.load(f)
return cls(data["tokens"])
def get_vocab(self):
return dict(self.token_to_id)
def get_vocab_size(self):
return self.vocab_size
# Backward compatibility alias
EdenTokenizer = COLMTokenizer
# =============================================================================
# BUILD
# =============================================================================
def build_default_tokenizer():
"""Build tokenizer from the default token definitions (no corpus)."""
return build_tokenizer()
def build_tokenizer(corpus_path=None, hf_dataset=None, hf_config=None,
hf_split="train", hf_field="text"):
"""Build tokenizer, optionally augmented with corpus-derived words.
Token order in the final vocab:
1. Special tokens (<pad>, <bos>, <eos>, <unk>)
2. Whitespace (space, newline, tab)
3. Base word tokens (top 100 English β€” lowercase)
4. Sentence-case base (The, Be, To, ... β€” +100)
5. Corpus freq words (lowercase)
6. Corpus freq words (Sentence case)
7. Corpus long words (lowercase)
8. Corpus long words (Sentence case)
9. Letters (A-Z, a-z)
10. Digits (0-9)
11. Symbols
"""
token_list = []
seen = set()
def add(tok):
if tok not in seen:
token_list.append(tok)
seen.add(tok)
# --- 1. Special tokens ---
for t in SPECIAL_TOKENS:
add(t)
# --- 2. Whitespace ---
for t in WHITESPACE_TOKENS:
add(t)
# --- 3. Base word tokens (lowercase) ---
for t in WORD_TOKENS:
add(t)
# --- 4. Sentence-case base words ---
for t in WORD_TOKENS:
add(make_sentence_case(t))
# --- 5-8. Corpus-derived words (if corpus provided) ---
corpus_words = []
has_corpus = (corpus_path is not None or hf_dataset is not None)
if has_corpus:
lines = iter_corpus_lines(
corpus_path=corpus_path,
hf_dataset=hf_dataset,
hf_config=hf_config,
hf_split=hf_split,
hf_field=hf_field,
)
counts = count_corpus_words(lines)
# Build set of existing lowercase words to exclude
existing = set(w.lower() for w in WORD_TOKENS)
freq_words, long_words, combined = select_corpus_words(
counts, existing, n_freq=50, n_long=50
)
corpus_words = combined
# Add corpus frequency words: lowercase then sentence case
for w in freq_words:
add(w)
for w in freq_words:
add(make_sentence_case(w))
# Add corpus long words: lowercase then sentence case
for w in long_words:
add(w)
for w in long_words:
add(make_sentence_case(w))
# --- 9. Letters ---
for t in UPPER:
add(t)
for t in LOWER:
add(t)
# --- 10. Digits ---
for t in DIGITS:
add(t)
# --- 11. Symbols ---
for t in sorted(SYMBOLS):
add(t)
tok = COLMTokenizer(token_list)
# Print breakdown
n_special = len(SPECIAL_TOKENS)
n_ws = len(WHITESPACE_TOKENS)
n_base_words = len(WORD_TOKENS)
n_base_sc = sum(1 for t in WORD_TOKENS if make_sentence_case(t) not in
set(WORD_TOKENS)) # non-overlapping sentence case
n_corpus = len(corpus_words) * 2 if corpus_words else 0 # lc + sc
n_letters = len(UPPER) + len(LOWER)
n_digits = len(DIGITS)
# Everything else is symbols
n_symbols = (tok.vocab_size - n_special - n_ws - n_base_words
- n_letters - n_digits)
# Subtract corpus + sentence-case base from the "symbols" remainder
n_symbols = n_symbols - n_base_sc - n_corpus
# Recalculate properly
accounted = (n_special + n_ws + n_base_words + n_letters + n_digits)
n_other = tok.vocab_size - accounted
print(f"\nBreakdown:")
print(f" Special: {n_special}")
print(f" Whitespace: {n_ws}")
print(f" Base words (lc): {n_base_words}")
if has_corpus:
print(f" Base words (Sc): +sentence case")
print(f" Corpus words (lc+Sc): {n_corpus}")
print(f" Letters: {n_letters}")
print(f" Digits: {n_digits}")
print(f" Other (Sc/symbols): {n_other}")
print(f" Total: {tok.vocab_size}")
return tok
# =============================================================================
# CLI
# =============================================================================
if __name__ == "__main__":
parser = argparse.ArgumentParser(
description="Build word+char hybrid tokenizer",
formatter_class=argparse.RawDescriptionHelpFormatter,
epilog="""
Examples:
# Base tokenizer only
python tokenizer_train.py --output colm_tokenizer.json
# Scan a local text file for extra vocab
python tokenizer_train.py --corpus training_data.txt --output colm_tokenizer.json
# Scan multiple files via glob
python tokenizer_train.py --corpus "data/*.txt" --output colm_tokenizer.json
# Use a HuggingFace dataset
python tokenizer_train.py --hf-dataset wikitext --hf-config wikitext-103-raw-v1 \\
--hf-split train --hf-field text --output colm_tokenizer.json
""")
parser.add_argument("--output", default="colm_tokenizer.json",
help="Output path (default: colm_tokenizer.json)")
parser.add_argument("--corpus", default=None,
help="Path or glob pattern for plain text file(s)")
parser.add_argument("--hf-dataset", default=None,
help="HuggingFace dataset name")
parser.add_argument("--hf-config", default=None,
help="HuggingFace dataset config")
parser.add_argument("--hf-split", default="train",
help="HuggingFace split (default: train)")
parser.add_argument("--hf-field", default="text",
help="HuggingFace text column (default: text)")
parser.add_argument("--test", default=None,
help="Test string to encode/decode")
args = parser.parse_args()
tok = build_tokenizer(
corpus_path=args.corpus,
hf_dataset=args.hf_dataset,
hf_config=args.hf_config,
hf_split=args.hf_split,
hf_field=args.hf_field,
)
tok.save(args.output)
# Quick roundtrip test
test = args.test or "The Machine God, in the beginning, with the Flesh God's design..."
encoded = tok.encode(test)
decoded = tok.decode(encoded)
print(f"\nTest: \"{test}\"")
print(f"IDs: {encoded}")
print(f"Back: \"{decoded}\"")
print(f"Match: {test == decoded}")