fracagnetta/tinystories.BPE8192

TinyStories — BPE8192 (pre-tokenized)

A pre-tokenized version of the TinyStories V2 (GPT-4) corpus, encoded with a byte-pair-encoding (BPE) tokenizer trained on the dataset itself with a target vocabulary size of 8192.

The token streams are stored as flat int32 NumPy arrays, ready to be memory-mapped and sliced into context windows for training small language models. Stories are separated by a single <eos> token (id 1).

Files

File	Description
tinystories.BPE8192.train.npy	Tokenized train split: 529M tokens as a 1-D int32 array (2.1 GB).
tinystories.BPE8192.valid.npy	Tokenized validation split: 5.3M tokens as a 1-D int32 array (21 MB).
tinystories.BPE8192.tokenizer.json	The trained 🤗 tokenizers BPE tokenizer (load with Tokenizer.from_file).
tinystories.BPE8192.meta.json	Metadata: vocab size, special-token ids, tokenizer filename, split sizes.

meta.json contents

{
  "vocab_size": 8192,
  "eos_token_id": 1,
  "unk_token_id": 0,
  "tokenizer_file": "tinystories/tinystories.BPE8192.tokenizer.json",
  "train_size": 529209187,
  "valid_size": 5343921
}

Tokenizer details

• Model: BPE with [UNK] as the unknown token.
• Pre-tokenizer: Whitespace (merges never cross whitespace/word boundaries).
• Special tokens: [UNK] (id 0), <eos> (id 1).
• Vocabulary size: 8192 (target; the trainer was given vocab_size=8192).

Each story was encoded independently and concatenated with a trailing <eos> (id 1) to delimit document boundaries in the flat token stream.

Usage

import json
import numpy as np
from tokenizers import Tokenizer
from huggingface_hub import hf_hub_download

REPO_ID = "fracagnetta/tinystories.BPE8192"

# --- metadata ---
meta_path = hf_hub_download(REPO_ID, "tinystories.BPE8192.meta.json", repo_type="dataset")
with open(meta_path) as f:
    meta = json.load(f)
print(meta)  # vocab_size, eos_token_id, unk_token_id, tokenizer_file, train/valid sizes

# --- tokenizer ---
tok_path = hf_hub_download(REPO_ID, "tinystories.BPE8192.tokenizer.json", repo_type="dataset")
tokenizer = Tokenizer.from_file(tok_path)

# --- token arrays (memory-mapped to avoid loading 2 GB into RAM) ---
train_path = hf_hub_download(REPO_ID, "tinystories.BPE8192.train.npy", repo_type="dataset")
valid_path = hf_hub_download(REPO_ID, "tinystories.BPE8192.valid.npy", repo_type="dataset")

train = np.load(train_path, mmap_mode="r")
valid = np.load(valid_path, mmap_mode="r")

print(len(train), len(valid))             # 529209187  5343921

# decode the first story of the validation split
print(tokenizer.decode(valid[:128].tolist()))

## Reproduction

The tokenizer is trained on the concatenated train+valid stories (split on `<|endoftext|>`),
then each split is encoded and saved as an `int32` `.npy` array with `<eos>` separators.

## License

Derived from [roneneldan/TinyStories](https://huggingface.co/datasets/roneneldan/TinyStories);
please follow the upstream dataset's license and terms.