gpt2-medium-indonesian / replace_token_script.py
alvin
update <|endoftext|> tokenizer id from 50257 to 50256
34c19b7
raw
history blame
2.93 kB
''''This script was used to replace the final index of tokenizer.json and vocab.json
with "<|endoftext|>" token. Also reassociate the corresponding merges'''
import json
tokenizer_path = 'tokenizer.json'
model_config_path = 'config.json'
vocab_path = 'vocab.json'
with open(vocab_path, "r") as f:
vocab_data = json.load(f)
with open(tokenizer_path, "r") as f:
tokenizer_data = json.load(f)
with open(model_config_path, "r") as f:
model_config = json.load(f)
model_vocab_size = model_config['vocab_size']
tokenizer_vocab = tokenizer_data['model']['vocab']
mergeslength = len(tokenizer_data['model']['merges'])
#readjust added_tokens 'id' to model_vocab_size - 1
tokenizer_data['added_tokens'][-1]['id'] = model_vocab_size - 1
final_index = model_vocab_size - 1
eos = '<|endoftext|>'
#retrieve the key of final index
old_key_final_index_tokenizer = list(tokenizer_data['model']['vocab'].keys())[final_index]
old_key_final_index_vocab = list(vocab_data.keys())[final_index]
old_key_final_index_vocab_min2 = list(vocab_data.keys())[final_index - 1]
old_key_final_index_tokenizer_merges = tokenizer_data['model']['merges'][mergeslength - 1]
print(f"old_key_final_index_tokenizer = {old_key_final_index_tokenizer}")
print(f"old_key_final_index_vocab = {old_key_final_index_vocab}")
print(f"old_key_final_index_vocab_min2 = {old_key_final_index_vocab_min2}")
print(f"old_key_final_index_tokenizer_merges = {old_key_final_index_tokenizer_merges}")
#replace old key with new key
tokenizer_data['model']['vocab']['<|endoftext|>'] = tokenizer_data['model']['vocab'][old_key_final_index_tokenizer]
vocab_data[eos] = vocab_data[old_key_final_index_vocab]
#replace the final merges idx with vocab_data - 1
tokenizer_data['model']['merges'] = tokenizer_data['model']['merges'][: mergeslength - 1]
#delete old key
del tokenizer_data['model']['vocab'][old_key_final_index_tokenizer]
del vocab_data[old_key_final_index_vocab]
#check updated key
old_key_final_index_tokenizer = list(tokenizer_data['model']['vocab'].keys())[final_index]
old_key_final_index_vocab = list(vocab_data.keys())[final_index]
old_key_final_index_tokenizer_merges = tokenizer_data['model']['merges'][mergeslength - 2]
print(len(tokenizer_data['model']['merges']))
print()
print(f"updated old_key_final_index_tokenizer = {old_key_final_index_tokenizer}")
print(f"updated old_key_final_index_vocab = {old_key_final_index_vocab}")
print(f"updated old_key_final_index_tokenizer_merges = {old_key_final_index_tokenizer_merges}")
with open(tokenizer_path, "w")as f:
json.dump(tokenizer_data, f)
with open(vocab_path, "w")as f:
json.dump(vocab_data, f)
with open('merges.txt') as f:
lines = f.readlines()
with open("merges.txt", "w") as f:
for i in range(len(lines) - 1):
f.write(lines[i])
with open('merges.txt') as f:
newlines = f.readlines()
print(f"newlines[len(newlines) - 1] = {newlines[len(newlines) - 1]}")