TinyStories: How Small Can Language Models Be and Still Speak Coherent English?

By Ronen Eldan, Yuan-Fang Li. arXiv.org, 2023

495 citations recorded by the source corpus.

Language models (LMs) are powerful tools for natural language processing, but they often struggle to produce coherent and fluent text when they are small. Models with around 125M parameters such as GPT-Neo (small) or GPT-2 (small) can rarely generate coherent and consistent English text beyond a few words even after extensive training. This raises the question of whether the emergence of the ability to produce coherent English text only occurs at larger scales (with hundreds of millions of parameters or more) and complex architectures (with many layers of global attention). In this work, we introduce TinyStories, a synthetic dataset of short stories that only contain words that a typical 3 to 4-year-olds usually understand, generated by GPT-3.5 and GPT-4. We show that TinyStories can be used to train and evaluate LMs that are much smaller than the state-of-the-art models (below 10 million total parameters), or have much simpler architectures (with only one transformer block), yet still produce fluent and consistent stories with several paragraphs that are diverse and have almost perfect grammar, and demonstrate reasoning capabilities. We also introduce a new paradigm for the evaluation of language models: We suggest a framework which uses GPT-4 to grade the content generated by these models as if those were stories written by students and graded by a (human) teacher. This new paradigm overcomes the flaws of standard benchmarks which often requires the model's output to be very structures, and moreover provides a multidimensional score for the model, providing scores for different capabilities such as grammar, creativity and consistency. We hope that TinyStories can facilitate the development, analysis and research of LMs, especially for low-resource or specialized domains, and shed light on the emergence of language capabilities in LMs.

Subfields: Artificial Intelligence

View Paper: https://www.semanticscholar.org/paper/28085f480ce456a376ebace9b899e3bc93dbc048

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

Citations in the corpus

The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A" (2023, unknown): “GPT-4 scores 28% on “Child”.”
CRAFT Your Dataset: Task-Specific Synthetic Dataset Generation Through Corpus Retrieval and Augmentation (2024, unknown): “Finally, we enhance dataset diversity by performing a deduplication step, filtering out generated samples that are highly similar to the seed few-shots or to each other using fuzzy string matching (Ranjith et al., 2022) with a token set similarity ratio of 0.85.”
Textbooks Are All You Need II: phi-1.5 technical report (2023, background): “The prior work [EL23] considered this question for the task of “speaking fluent English”, while the subsequent work [GZA + 23] considered the more challenging task of coding simple functions in Python.”
MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Models (2023, background): “This is also aligned with the findings with TinyStories [18], where short and easy story data can help LLMs generate content fluently.”
On LLMs-Driven Synthetic Data Generation, Curation, and Evaluation: A Survey (2024, unknown): “…and testing data, LLM-driven synthetic data generation also may pave the way for developing next-generation LLMs. Insights from TinySto-ries (Eldan and Li, 2023) and the Phi series (Gu-nasekar et al., 2023; Li et al., 2023b) emphasize that data quality is crucial for effective model…”
Textbooks Are All You Need (2023, methodology): “To evaluate candidate solutions, we therefore adopt the approach of using GPT-4 to grade the solution (such as in [EL23]).”
Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2 (2024, unknown): “Braun et al. (2024) trained different variants of SAEs on GPT-2 small (Radford et al., 2019) and Tinystories-1M (Eldan and Li, 2023) on the residual stream activations in select layers.”
Sparse Prototype Code Underlies Classification and Prediction Across Modalities (2026, unknown): “…from pretrained backbones spanning vision (ImageNet-1K [Deng et al., 2009]), language (masked and next-token prediction on TinyStories [Eldan and Li, 2023]), and audio (VGGSound [Chen et al., 2020]), across transformer and convolutional architectures, supervised, self-supervised, and…”