Chinchilla: Training Compute-Optimal Large Language Models


Paper: Hoffmann, J., et al. (DeepMind), Training Compute-Optimal Large Language Models, 2022. arXiv:2203.15556

Why this paper matters

Kaplan’s scaling laws (previous post) implied “make the model as big as possible and feed it whatever data you have.” DeepMind re-ran the analysis with far more rigor — over 400 training runs — and found the field had been wasting compute. Under a fixed compute budget, parameters and training tokens should grow in equal proportion, not lopsidedly toward size. The 70B Chinchilla, trained on 1.4T tokens, beat Gopher (280B) and was competitive with models 4× its size. If you set a training budget today, this paper is the recipe you use.

The core idea: equal scaling

Write the loss as a sum of power-law terms in parameters $N$ and data $D$:

\[L(N, D) = \frac{A}{N^{\alpha}} + \frac{B}{D^{\beta}} + E\]

Fitting to the 400+ runs gives $\alpha \approx 0.34$, $\beta \approx 0.28$ (much larger exponents than Kaplan’s — scaling helps more than previously thought). Minimizing under a compute constraint $C \approx 6 N D$ yields:

\[N_{\text{opt}} \propto C^{0.5}, \qquad D_{\text{opt}} \propto C^{0.5}\]

The practical takeaway, often quoted as a rule of thumb: train on roughly 20× as many tokens as parameters ($D \approx 20 N$).

What this corrected

At the time, the flagship models were off the optimum:

Chinchilla showed you could match or beat them with a fraction of the parameters if you simply fed them more data.

Key results

Why it matters today

Every serious training run since 2022 sizes its corpus from Chinchilla math, not Kaplan’s. The lesson generalizes beyond text: the same 1:20 intuition shows up (with caveats) in vision and multimodal models. Pair this with the Scaling Laws post — together they are the “how big / how much data” chapter of any LLM course.

References

© 2026 Jiawei    粤ICP备18035774号