TinyStories: How Small Can Language Models Be and Still Speak Coherent English?
Paper • 2305.07759 • Published • 48
Dataset Viewer
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
CS 5236: Advanced Generative AI and Agents
Programming Assignment 1: The Modern Transformer LM
This dataset accompanies Programming Assignment 1 for CS 5236: Advanced Generative AI and Agents. It gives every student the same ready-to-use text corpus and tokenizer for implementing and training a modern Transformer language model from scratch.
Source and credits
The text comes from roneneldan/TinyStories,
introduced by Ronen Eldan and Yuanzhi Li in
TinyStories: How Small Can Language Models Be and Still Speak Coherent English?.
Tokenization uses the 8,192-token byte-level BPE tokenizer published with
commonsense-ai/tinystories-15m.
We gratefully acknowledge the authors and maintainers of both resources.
What this repository contains
Each .bin file is a flat stream of token IDs. Every story is followed by one
<|endoftext|> token. The streams do not contain padding or pre-built training
windows, so your batch sampler can choose its own context length.
| File | Purpose | Documents | Tokens |
|---|---|---|---|
data/train.bin |
Complete training split | 2,119,719 | 466,876,982 |
data/validation.bin |
Complete validation split | 21,990 | 4,692,376 |
debug/train.bin |
Small development stream | 926 | 200,116 |
debug/validation.bin |
Small development stream | 231 | 50,208 |
tokenizer/tokenizer.json |
Fixed course tokenizer | -- | -- |
The supplied tokenizer/tokenizer.json is used to encode generation prompts and
decode model outputs. The Transformer embeddings and language-model head are
still randomly initialized and trained by you.
- Downloads last month
- 51