jdpressman/comma_v0.1_training_dataset_sample_10B

Comma v0.1 Training Dataset (10 Billion Token Sample)

This is a 10 billion token subset of the Comma v0.1 Training Set intended as a convenience for small deep learning experiments. It is similar in spirit to the 1 billion token RedPajama sample which is no longer functioning with HuggingFace transformers due to involving the execution of arbitrary code at load time.

Method

The data was subsetted using the following script:

import os
import json
import gzip
import math
import random
import requests
from pathlib import Path
from tqdm import tqdm
from datasets import load_dataset
from transformers import AutoTokenizer
from argparse import ArgumentParser

parser = ArgumentParser()
parser.add_argument("--output-dir", type=Path, default="shards")
parser.add_argument("--tokens", default=10**9, type=int,
                    help="The number of tokens to subset.")
parser.add_argument("--shard-size", type=int, default=(250 * (10 ** 6)))
args = parser.parse_args()

tokenizer = AutoTokenizer.from_pretrained("common-pile/comma-v0.1-1t")
dataset = load_dataset("common-pile/comma_v0.1_training_dataset")

if not os.path.exists("shards"):
    os.mkdir("shards")

used = set()
token_count = 0
shard_index = 0
if os.path.exists("subset_resume.json"):
    with open("subset_resume.json") as infile:
        data = json.load(infile)
        spans = set(data["used"])
        token_count = data["token_count"]
        shard_index = data["shard_index"]

num_shards = math.ceil(args.tokens / args.shard_size)
milestone = args.shard_size
progress = tqdm(total=args.tokens)
while token_count < args.tokens:
    progress.set_description(f"Tokens Processed (Shard {shard_index})")
    filename = f"train-{shard_index:05d}-of-{num_shards:05d}.jsonl.gz"
    filepath = Path(args.output_dir) / filename
    with gzip.open(filepath, 'wt', encoding='utf-8') as outfile:
        while token_count < milestone:
            choices = set()
            for i in range(64):
                choice = random.randrange(dataset["train"].num_rows)
                while choice in used:
                    choice = random.randrange(dataset["train"].num_rows)
                used.add(choice)
                choices.add(choice)
            assert len(choices) == 64
            items = []
            for choice in choices:
                items.append(dataset["train"][choice])
            texts = [item["text"] for item in items]
            new_tokens = sum([len(i) for i in tokenizer(texts)["input_ids"]])
            token_count += new_tokens
            progress.update(new_tokens)
            for item in items:
                json_line = json.dumps(item)
                outfile.write(json_line + "\n")
            if token_count > milestone:
                with open("subset_resume.json", "w") as outfile:
                    serial_used = list(used)
                    json.dump({"used":serial_used, "token_count":token_count, "shard_index":shard_index}, outfile)
        milestone += args.shard_size
    shard_index += 1

Feel free to adapt and use this script to make other subsets.