File size: 9,045 Bytes
d35dd87 | 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 | """Explicit biological BPE -> model ID mapping; natural language uses base BPE.
Source tokenizers are the historical OmniGene BPEs, with missing single-character
alphabet entries appended without altering existing IDs or merges. Every ordinary
source piece (including single characters) receives a separate modality-specific
model ID. Token display names are serialization metadata, never input text.
"""
from __future__ import annotations
import argparse
import hashlib
import json
from pathlib import Path
import shutil
from tokenizers import AddedToken, Tokenizer
from transformers import AutoTokenizer
ROOT = Path(__file__).resolve().parents[1]
DEFAULT = ROOT / 'artifacts/laya_direct_bpe_legacy_representation'
LEGACY = Path('/root/autodl-fs/omnigene_v2/scripts/vocab/trained_bpe')
def digest(path):
return hashlib.sha256(Path(path).read_bytes()).hexdigest()
def prepare(output=DEFAULT, source_root=LEGACY):
output = Path(output)
if output.exists():
raise FileExistsError(f'Refusing to overwrite representation: {output}')
source_root = Path(source_root)
source_names = {'dna': 'dna_bpe_20k.json', 'protein': 'protein_bpe_8k.json'}
repaired_sources, repairs = {}, {}
for kind,name in source_names.items():
original = json.loads((source_root/name).read_text())
vocab = original['model']['vocab']
# The original protein BPE silently drops N; DNA similarly lacks N.
# Preserve all learned merges and append only a fixed biological alphabet.
alphabet = 'ACGTN' if kind=='dna' else 'ABCDEFGHIJKLMNOPQRSTUVWXYZ'
missing = [letter for letter in alphabet if letter not in vocab]
for letter in missing:
vocab[letter] = max(vocab.values())+1
repairs[kind] = missing
repaired_sources[kind] = Tokenizer.from_str(json.dumps(original))
base = AutoTokenizer.from_pretrained(ROOT / 'artifacts/laya_model/tokenizer')
expanded = AutoTokenizer.from_pretrained(ROOT / 'artifacts/laya_model/tokenizer')
entries = []
for kind, name in source_names.items():
source = repaired_sources[kind]
for piece, source_id in sorted(source.get_vocab().items(), key=lambda pair: pair[1]):
if piece in {'[UNK]', '[PAD]'}:
continue
if not piece.isalpha():
raise ValueError(f'Unexpected source vocabulary piece: {piece!r}')
ids = base(piece, add_special_tokens=False)['input_ids']
if not ids or any(i in base.all_special_ids for i in ids):
raise ValueError(f'Cannot initialize {kind}:{piece} from raw fragment')
entries.append({'modality': kind, 'source_id': source_id, 'piece': piece,
'token': f'<BIO_{kind.upper()}:{piece}>', 'base_ids': ids})
added = expanded.add_tokens([AddedToken(x['token'], normalized=False) for x in entries])
if added != len(entries):
raise ValueError('Duplicate or conflicting vocabulary entries')
vocab = expanded.get_vocab()
for entry in entries:
entry['expanded_id'] = vocab[entry['token']]
if any(vocab[k] != v for k, v in base.get_vocab().items()):
raise ValueError('Original token IDs changed')
output.mkdir(parents=True)
(output/'source_tokenizers').mkdir()
(output/'original_source_tokenizers').mkdir()
for kind,name in source_names.items():
repaired_sources[kind].save(str(output/'source_tokenizers'/name))
shutil.copy2(source_root/name, output/'original_source_tokenizers'/name)
scripts = source_root.parent
provenance_scripts = [scripts/name for name in ('1-sample_bpe_corpus.py','2-train_bio_bpe.py','3-expand_tokenizer.py')]
(output/'provenance_scripts').mkdir()
for path in provenance_scripts:
if path.exists():shutil.copy2(path,output/'provenance_scripts'/path.name)
base.save_pretrained(output / 'base_tokenizer')
expanded.save_pretrained(output / 'expanded_tokenizer')
(output / 'new_tokens.json').write_text(json.dumps(entries, ensure_ascii=False, indent=2)+'\n')
meta = {'schema_version': 'laya-direct-bpe-legacy-v1', 'no_cpt': True,
'fit_policy': 'reuse external historical BPE; no refitting; no neural CPT',
'historical_fit_recipe': 'sample about 1 GiB each from dna_32g.txt and protein_uni_16.txt, seed42; BPE min_frequency10',
'historical_fit_corpus_overlap_with_evaluation': 'not established; no train-only claim',
'source_files': {k:{'path':str(source_root/n),'sha256':digest(source_root/n)} for k,n in source_names.items()},
'alphabet_repair': repairs, 'existing_merges_preserved':True,
'old_vocab_size': len(base), 'new_vocab_size': len(expanded),
'added_by_modality': {k: sum(x['modality'] == k for x in entries) for k in source_names},
'initialization': 'mean of original tokenizer IDs of RAW biological fragment; no delimiters',
'encoding': 'source BPE ID mapped directly to added model ID; natural language uses base tokenizer',
'output_hashes': {str(p.relative_to(output)): digest(p) for p in sorted(output.rglob('*')) if p.is_file()}}
(output / 'metadata.json').write_text(json.dumps(meta, indent=2)+'\n')
print(json.dumps({k: meta[k] for k in ('old_vocab_size','new_vocab_size','added_by_modality')}, indent=2))
class DirectBPE:
def __init__(self, directory=DEFAULT):
self.root = Path(directory)
self.meta = json.loads((self.root / 'metadata.json').read_text())
for name, expected in self.meta['output_hashes'].items():
if digest(self.root / name) != expected:
raise ValueError(f'Representation hash mismatch: {name}')
self.base = AutoTokenizer.from_pretrained(self.root / 'base_tokenizer')
self.expanded = AutoTokenizer.from_pretrained(self.root / 'expanded_tokenizer')
self.entries = json.loads((self.root / 'new_tokens.json').read_text())
self.sources = {k: Tokenizer.from_file(str(self.root / 'source_tokenizers' / name))
for k, name in [('dna','dna_bpe_20k.json'),('protein','protein_bpe_8k.json')]}
self.maps = {k: {e['source_id']: e['expanded_id'] for e in self.entries if e['modality'] == k}
for k in self.sources}
self.inverse = {e['expanded_id']: e['piece'] for e in self.entries}
def sequence_ids(self, sequence, kind):
encoded = self.sources[kind].encode(sequence)
if ''.join(encoded.tokens) != sequence:
raise ValueError('Source BPE changed the sequence or produced UNK')
try:
ids = [self.maps[kind][i] for i in encoded.ids]
except KeyError as exc:
raise ValueError('Unmapped source BPE token') from exc
if ''.join(self.inverse[i] for i in ids) != sequence:
raise ValueError('Model IDs do not reconstruct the original sequence')
return ids
def build(self, row, condition, builder, order, max_len=1024, head_max_len=256):
q = {'t': 'choice', 'ins': 'Choose the correct biological label for the sequence.',
'crit': {x: None for x in row['choices']}}
prefix = f"Task context: {row['context']}\nSequence: "
if condition == 'raw':
ids, markers = builder(self.base, prefix + row['sequence'], q,
max_len=1_000_000, head_max_len=head_max_len, option_order=order)
elif condition == 'full_bpe':
# Reuse the official header exactly, including candidate markers.
# Its final SEP terminates the empty state; replace that with our
# explicitly tokenized state and restore the final SEP.
header, markers = builder(self.base, '', q, max_len=1_000_000,
head_max_len=head_max_len, option_order=order)
context_ids = self.base(prefix.replace(self.base.mask_token, ' '), add_special_tokens=False)['input_ids']
ids = header[:-1] + context_ids + self.sequence_ids(row['sequence'], row['kind']) + [self.base.sep_token_id]
else:
raise ValueError(condition)
if len(ids) > max_len:
raise ValueError(f"Full input exceeds {max_len}: {row['id']} {condition} {len(ids)}")
if len(markers) != len(row['choices']):
raise ValueError('Candidate marker loss')
for canonical, marker in zip(order, markers):
expected = self.base(' '+row['choices'][canonical], add_special_tokens=False)['input_ids']
if ids[marker+1:marker+1+len(expected)] != expected:
raise ValueError('Candidate text truncated or remapped incorrectly')
return ids, markers
if __name__ == '__main__':
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('--output', type=Path, default=DEFAULT)
parser.add_argument('--source-dir', type=Path, default=LEGACY)
args=parser.parse_args()
prepare(args.output,args.source_dir)
|