victorhu493/cs336-basics-collection

CS336 Assignment 1 — Pre-tokenized Data & BPE Tokenizers

This repository contains preprocessing artifacts produced for Stanford CS336: Language Modeling from Scratch, Spring 2025 — Assignment 1: Basics.

It includes:

• pre-tokenized TinyStories train/validation data,
• pre-tokenized OpenWebText (OWT sample) train/validation data,
• byte-level BPE vocabularies and merge tables for both datasets.

The main purpose of this repository is to avoid repeating the relatively expensive tokenization step when running language-model training experiments.

Note: This is a personal course-project artifact and is not an official Stanford CS336 release.

Background

CS336 Assignment 1 builds a Transformer language model from scratch, including:

1. a byte-level BPE tokenizer,
2. Transformer model components,
3. cross-entropy loss and AdamW,
4. checkpointing and a training loop,
5. language-model training and evaluation on TinyStories and OpenWebText.

The Spring 2025 assignment uses two tokenizer settings:

Dataset	BPE vocabulary size
TinyStories	10,000
OpenWebText	32,000

The special token <|endoftext|> is included in the tokenization setup.

Repository Structure

cs336-basics-collection/
├── tokenized/
│   ├── tinystories_train_ids.npy
│   ├── tinystories_valid_ids.npy
│   ├── owt_train_ids.npy
│   └── owt_valid_ids.npy
│
└── tokenizer/
    ├── tinystories/
    │   ├── vocab.json
    │   └── merges.json
    │
    └── owt/
        ├── vocab.json
        └── merges.json

Tokenized data

The .npy files contain contiguous sequences of integer token IDs obtained by encoding the corresponding train or validation corpus with its dataset-specific BPE tokenizer.

They are intended for causal language-model training, where random contiguous windows can be sampled directly from the token array.

File	Description
tokenized/tinystories_train_ids.npy	TinyStories training tokens
tokenized/tinystories_valid_ids.npy	TinyStories validation tokens
tokenized/owt_train_ids.npy	OpenWebText training tokens
tokenized/owt_valid_ids.npy	OpenWebText validation tokens

Tokenizers

Each tokenizer directory contains:

• vocab.json: mapping between token IDs and byte/token representations
• merges.json: learned BPE merge operations

The token IDs in the .npy files must be interpreted with the corresponding tokenizer.

In particular, TinyStories and OpenWebText use different vocabularies, so their token IDs are not interchangeable.

Usage

Install the required packages:

pip install numpy huggingface_hub

Download a tokenized split:

from huggingface_hub import hf_hub_download
import numpy as np

path = hf_hub_download(
    repo_id="victorhu493/cs336-basics-collection",
    repo_type="dataset",
    filename="tokenized/tinystories_train_ids.npy",
)

tokens = np.load(path, mmap_mode="r")

print(tokens.shape)
print(tokens.dtype)
print(tokens[:20])

Using mmap_mode="r" is recommended for the larger arrays, particularly OpenWebText, since the full dataset does not need to be loaded into RAM.

Sampling a language-model batch

A standard next-token prediction batch can be constructed directly from the token array:

import numpy as np
import torch

def get_batch(
    tokens,
    batch_size=32,
    context_length=256,
):
    starts = np.random.randint(
        0,
        len(tokens) - context_length - 1,
        size=batch_size,
    )

    x = np.stack([
        tokens[i : i + context_length]
        for i in starts
    ])

    y = np.stack([
        tokens[i + 1 : i + context_length + 1]
        for i in starts
    ])

    return (
        torch.from_numpy(x.astype(np.int64)),
        torch.from_numpy(y.astype(np.int64)),
    )

This produces:

x: [batch_size, context_length]
y: [batch_size, context_length]

where y is x shifted by one token, as required for autoregressive language modeling.

Original Data

The preprocessing follows the datasets used in the Spring 2025 CS336 Assignment 1 setup:

• TinyStories — roneneldan/TinyStories
• OpenWebText sample — stanford-cs336/owt-sample

The original CS336 assignment downloads the raw TinyStories and OpenWebText text files and then asks students to train BPE tokenizers and encode the corpora into integer token sequences.

This repository contains the resulting derived tokenizer and token-ID artifacts, rather than the original raw text files.

Important Notes

Tokenizer compatibility

Always pair each tokenized dataset with its corresponding tokenizer:

tinystories_*  <-> tokenizer/tinystories/
owt_*          <-> tokenizer/owt/

Using the wrong vocabulary will assign different meanings to the same integer IDs.

Hugging Face Dataset Viewer

These files are stored as NumPy token arrays rather than row-oriented Parquet/Arrow data.

Therefore, this repository is primarily intended to be consumed directly with NumPy / PyTorch rather than through the Hugging Face Dataset Viewer.

Reproducibility

These artifacts are useful when reproducing CS336-style language-model training experiments without retraining the BPE tokenizer or re-encoding the entire corpus.

For exact reproduction, model hyperparameters, random seeds, batching strategy, optimizer settings, and training code should still be recorded separately.

References

• Stanford CS336 — Language Modeling from Scratch, Spring 2025
• stanford-cs336/assignment1-basics
• stanford-cs336/owt-sample
• roneneldan/TinyStories

Disclaimer

This repository is an independent educational artifact created while working through CS336.

It is not affiliated with, endorsed by, or maintained by Stanford University or the CS336 course staff.

Please refer to the original datasets and course materials for their respective licensing and usage terms.