train_sanghara_multilingual_tokenizer.py

from datasets import concatenate_datasets, load_dataset, load_from_disk
import argparse
from tokenizers import Tokenizer, decoders, models, pre_tokenizers, processors, trainers
from transformers import GPT2TokenizerFast, AutoTokenizer
from datasets import config
from datasets import DatasetDict, Dataset
import logging

def initialize_logger(log_file):
    logging.basicConfig(filename=log_file, level=logging.INFO, format='%(asctime)s: %(message)s')

def log_parameters(vocab_size, batch_size, fertility_score, proportion_continued_words, log_file='parameters.log'):
    initialize_logger(log_file)
    logging.info(f"Vocabulary Size: {vocab_size}, Batch Size: {batch_size}, Fertility Score: {fertility_score}, Proportion of Continued word: {proportion_continued_words}")

def parse_arguments():
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--batch_size",
        type=int,
        required=True,
        help="Batch size to use for training"
    )
    parser.add_argument(
        "--vocab_size",
        type=int,
        required=True,
        help="Vocabulary size to use for tokenizer"
    )
    parser.add_argument(
        "--use_config",
        choices=['xlm-roberta', 'vanilla','llama'],
        required=True,
        help="Use XLM-RoBERTa config or Vanilla BPE"
    )
    parser.add_argument(
        "--do_evaluate",
        action='store_true',
        help="Enable evaluation."
    )
    args = parser.parse_known_args()
    return args
def train_tokenizer(args):

    code_dataset_go= load_dataset('code_x_glue_ct_code_to_text','go',split='train', cache_dir='/sml2/atul/CENTRAL_CACHE')['code']
    code_dataset_java= load_dataset('code_x_glue_ct_code_to_text','java',split='train', cache_dir='/sml2/atul/CENTRAL_CACHE')['code']
    code_dataset_javascript= load_dataset('code_x_glue_ct_code_to_text','javascript',split='train', cache_dir='/sml2/atul/CENTRAL_CACHE')['code']
    code_dataset_php= load_dataset('code_x_glue_ct_code_to_text','php',split='train', cache_dir='/sml2/atul/CENTRAL_CACHE')['code']
    code_dataset_python= load_dataset('code_x_glue_ct_code_to_text','python',split='train', cache_dir='/sml2/atul/CENTRAL_CACHE')['code']
    code_dataset_ruby= load_dataset('code_x_glue_ct_code_to_text','ruby',split='train', cache_dir='/sml2/atul/CENTRAL_CACHE')['code']

    indic_datasets_hi= load_dataset('ai4bharat/sangraha', data_dir="verified/hin", cache_dir='/sml2/atul/CENTRAL_CACHE')['train']['text'][:6000000]
    indic_datasets_bn= load_dataset('ai4bharat/sangraha', data_dir="verified/ben", cache_dir='/sml2/atul/CENTRAL_CACHE')['train']['text'][:6000000]
    wikipedia_en = load_dataset("wikipedia", "20220301.en", cache_dir='/sml2/atul/CENTRAL_CACHE')['train']['text'][:1000000]
    

    combined_train_set=code_dataset_go+code_dataset_java+code_dataset_javascript+code_dataset_php+code_dataset_python+code_dataset_ruby+indic_datasets_hi+indic_datasets_bn+wikipedia_en

    data = {
            "train":{"text": combined_train_set},
            "validation": {"text": []},
            "test": {"text": []},
        }
        # print(data)
    custom_dataset = DatasetDict()
    for split in data:
        custom_dataset[split] = Dataset.from_dict(data[split])
    custom_dataset=custom_dataset["train"]

    def batch_iterator():
        for idx in range(0, len(custom_dataset), args.batch_size):
            yield custom_dataset[idx: idx + args.batch_size]['text']

    new_line="\n"
    replacing_dict={}
    for i in range(5,25):
        replacecable_token="<|reserved_special_token_"+str(i)+"|>"
        replacing_dict[replacecable_token]=new_line
        new_line+="\n"
    if args.use_config == 'vanilla':
        tokenizer = Tokenizer(models.BPE())
        tokenizer.pre_tokenizer = pre_tokenizers.ByteLevel(add_prefix_space=False)
        print(f"[INFO] The brown fox jumped over the lazy dog\n{tokenizer.pre_tokenizer.pre_tokenize_str('The brown fox jumped over the lazy dog')}")
        print(f"[INFO] Training...")
        trainer = trainers.BpeTrainer(vocab_size=args.vocab_size, special_tokens=["<|endoftext|>"])
        tokenizer.train_from_iterator(batch_iterator(), trainer=trainer)
        tokenizer.post_processor = processors.ByteLevel(trim_offsets=False)
        tokenizer.decoder = decoders.ByteLevel()
        tokenizer = GPT2TokenizerFast(tokenizer_object=tokenizer)
    elif args.use_config == 'xlm-roberta':
        tokenizer = AutoTokenizer.from_pretrained('xlm-roberta-base')
        trained_tokenizer = tokenizer.train_new_from_iterator(batch_iterator(), vocab_size=args.vocab_size)
    elif args.use_config == 'llama':
        print("skipped")
        tokenizer = AutoTokenizer.from_pretrained('meta-llama/Meta-Llama-3-8B')
        print(tokenizer)
        trained_tokenizer = tokenizer.train_new_from_iterator(batch_iterator(), vocab_size=args.vocab_size, new_special_tokens=["<unk>","<pad>","<mask>"],special_tokens_map=replacing_dict)

    trained_tokenizer.save_pretrained('hi-sanghara-xlmr-bgpt-bpe-tokenizer1')
    print(f"[INFO] Tokenizer saved to disk")


def main():
    args, _ = parse_arguments()
    train_tokenizer(args)

if __name__ == "__main__":
    main()