Papers · Chinchilla
AI PAPER

Training Compute-Optimal Large Language Models

Proved that most large language models were dramatically undertrained, showing that smaller models trained on more data outperform larger models trained on less data for the same compute budget.

Authors
Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch +7
Institution
DeepMind
Published
March NaN, 2022
Citations
5,000
Category
Scaling
Impact
major
PAPER EXPLAINED

Plain-English Summary

The AI field had been racing to build ever-larger models, assuming bigger was always better. Chinchilla showed this was wrong — or at least inefficient. For a given compute budget, there is an optimal balance between model size and the amount of training data. Most existing models (including GPT-3) were too large for the amount of data they were trained on.

DeepMind proved this by training a 70-billion parameter model (Chinchilla) on 4x more data than the 280-billion parameter Gopher. Despite being much smaller, Chinchilla outperformed Gopher on nearly every benchmark. The lesson: you should scale model size and training data equally. Double your compute? Make the model 1.4x bigger AND train on 1.4x more data.

Think of it like cooking: using a massive oven (big model) but only cooking for five minutes (little data) wastes energy. A smaller oven running for the right amount of time produces a better meal for the same energy cost.

Key Innovation

The paper established revised scaling laws showing that model parameters and training tokens should be scaled in roughly equal proportion. Specifically, for compute-optimal training, the number of training tokens should scale linearly with model parameters. This contradicted the previous Kaplan et al. scaling laws, which suggested model size should grow faster than dataset size.

Chinchilla’s 70B parameters trained on 1.4 trillion tokens outperformed Gopher (280B parameters, 300 billion tokens), Megatron-Turing NLG (530B), and GPT-3 (175B) — all models that were “over-parameterized and under-trained” by Chinchilla’s analysis.

Impact on the Field

Chinchilla fundamentally redirected the AI scaling race. Instead of building the largest possible model, labs shifted to training appropriately-sized models on much more data. This directly influenced Meta’s LLaMA (trained on 1.4T tokens) and the entire open-source model movement, which realized that smaller, well-trained models could compete with much larger ones.

The paper also increased demand for training data, sparking legal and ethical debates about data sourcing that continue today. When data becomes the bottleneck rather than model size, questions about copyright, licensing, and synthetic data generation become urgent.

Models That Built on This

Meta’s LLaMA family followed Chinchilla’s principles directly, training relatively small models on massive datasets. Mistral’s efficient models embody the same philosophy. Google’s Gemini models were trained with Chinchilla-optimal ratios in mind. The paper effectively ended the “parameter count wars” and started the “training efficiency wars,” where labs compete on how much capability they can extract from each unit of compute.