āļ„āļđāđˆāļĄāļ·āļ­āļžāļ·āđ‰āļ™āļāļēāļ™

āļāļēāļĢāļāļķāļāļ­āļšāļĢāļĄāđ€āļžāļ·āđˆāļ­āļ›āļĢāļ°āļŠāļīāļ—āļ˜āļīāļ āļēāļžāļāļēāļĢāļ›āļĢāļ°āļĄāļ§āļĨāļœāļĨāļ‚āļ­āļ‡ Chinchilla

The Chinchilla paper estimates training-compute allocations with three complementary methods: training-curve envelopes, IsoFLOP profiles, and a parametric loss fit.

āļŸāļąāļ‡āļ„āļģāđāļ™āļ°āļ™āļģāļ™āļĩāđ‰āļšāļĢāļĢāļĒāļēāļĒāđ‚āļ”āļĒ AI · ~2 min

āļšāļ™āļŦāļ™āđ‰āļēāļ™āļĩāđ‰āļ­āđˆāļēāļ™ 2 āļ™āļēāļ—āļĩ
  1. āļ āļēāļžāļĢāļ§āļĄ
  2. āđ€āļˆāļēāļ°āļĨāļķāļ
  3. āļāļēāļĢāđƒāļŠāđ‰āļ‡āļēāļ™āļˆāļĢāļīāļ‡āđƒāļ™āđ‚āļĨāļāđāļŦāđˆāļ‡āļ„āļ§āļēāļĄāđ€āļ›āđ‡āļ™āļˆāļĢāļīāļ‡
  4. āļ­āļ™āļēāļ„āļ•āļ‚āļ­āļ‡āļāļēāļĢāļāļķāļāļ­āļšāļĢāļĄāļ—āļĩāđˆāđ€āļžāļīāđˆāļĄāļ›āļĢāļ°āļŠāļīāļ—āļ˜āļīāļ āļēāļžāļāļēāļĢāļ›āļĢāļ°āļĄāļ§āļĨāļœāļĨāļ‚āļ­āļ‡ Chinchilla
  5. āļāļēāļĢāļ™āļģāđ„āļ›āļ›āļāļīāļšāļąāļ•āļī
  6. āļŠāļģāļĢāļ§āļˆāļ•āđˆāļ­āđ„āļ›
  7. āļ„āļģāļ–āļēāļĄāļ—āļĩāđˆāļžāļšāļšāđˆāļ­āļĒ

āļ āļēāļžāļĢāļ§āļĄ

It matters because its measured range and estimation methods show how researchers can cross-check a compute-optimal forecast rather than rely on one extrapolation.

āđ€āļˆāļēāļ°āļĨāļķāļ

Hoffmann and colleagues trained over 400 language models, from about 70 million to over 16 billion parameters, on roughly 5 billion to 500 billion tokens. They used three approaches to estimate how model size and training tokens should share a compute budget. First, they fixed model sizes, varied training horizons, smoothed and interpolated training curves, and selected the lowest loss for a given FLOP count. Second, they constructed IsoFLOP profiles by varying model size at fixed compute budgets and fitting the loss minimum on each profile. Third, they fit a parametric loss function over parameter count and token count, using Huber loss and L-BFGS with a grid of initializations. The three methods produced similar near-equal scaling estimates in the studied regime. Chinchilla itself was a 70B model trained on 1.4T tokens using MassiveText, with a slightly different subset distribution from Gopher.

āļāļēāļĢāđƒāļŠāđ‰āļ‡āļēāļ™āļˆāļĢāļīāļ‡āđƒāļ™āđ‚āļĨāļāđāļŦāđˆāļ‡āļ„āļ§āļēāļĄāđ€āļ›āđ‡āļ™āļˆāļĢāļīāļ‡

Hold several model sizes fixed and vary training duration to estimate minimum loss at different compute budgets

Build IsoFLOP profiles by comparing model sizes at fixed compute budgets and locating each loss minimum

Fit a parametric loss function of model size and training tokens, then compare its frontier with curve-based estimates

Check a forecast against the data and training recipe used for the 70B Chinchilla model

āļ­āļ™āļēāļ„āļ•āļ‚āļ­āļ‡āļāļēāļĢāļāļķāļāļ­āļšāļĢāļĄāļ—āļĩāđˆāđ€āļžāļīāđˆāļĄāļ›āļĢāļ°āļŠāļīāļ—āļ˜āļīāļ āļēāļžāļāļēāļĢāļ›āļĢāļ°āļĄāļ§āļĨāļœāļĨāļ‚āļ­āļ‡ Chinchilla

Scaling studies can become more informative when they compare several estimation methods, vary dataset mixtures, and test their predictions on separately trained models. Forecasts beyond the measured parameter and token ranges remain uncertain. Future work should report the sweep design, fit assumptions, and evaluation data alongside any compute-optimal recommendation.

āļāļēāļĢāļ™āļģāđ„āļ›āļ›āļāļīāļšāļąāļ•āļī

āļ—āļģāđ„āļĄāļĄāļąāļ™āļ–āļķāļ‡āļŠāļģāļ„āļąāļ

  • āļāļēāļĢāļ•āļąāļ”āļŠāļīāļ™āđƒāļˆāļ—āļĩāđˆāļŠāļąāļ”āđ€āļˆāļ™āļĒāļīāđˆāļ‡āļ‚āļķāđ‰āļ™. āļŠāđˆāļ§āļĒāđƒāļŦāđ‰āļ„āļļāļ“āđāļĒāļāļ„āļģāļāļĨāđˆāļēāļ§āļ­āđ‰āļēāļ‡āļ—āļēāļ‡āđ€āļ—āļ„āļ™āļīāļ„āļ—āļĩāđˆāļŠāļąāļ”āđ€āļˆāļ™āļ­āļ­āļāļˆāļēāļāļ āļēāļĐāļēāļ—āļēāļ‡āļāļēāļĢāļ•āļĨāļēāļ”āđ„āļ”āđ‰
  • āļ•āđ‰āļ™āļ—āļļāļ™āđāļĨāļ°āļ‡āļšāļ›āļĢāļ°āļĄāļēāļ“. āļ„āļļāļ“āļŠāļēāļĄāļēāļĢāļ–āļ–āļēāļĄāļ„āļģāļ–āļēāļĄāļāļēāļĢāđƒāļŠāđ‰āļ‡āļēāļ™āļ—āļĩāđˆāļ”āļĩāļ‚āļķāđ‰āļ™āļāđˆāļ­āļ™āļ—āļĩāđˆāļˆāļ°āđƒāļŠāđ‰āđ€āļ‡āļīāļ™āļŦāļĢāļ·āļ­āđ€āļ§āļĨāļē
  • āļ—āļĩāļĄāļ‡āļēāļ™āđāļĨāļ°āļ‚āļąāđ‰āļ™āļ•āļ­āļ™āļāļēāļĢāļ—āļģāļ‡āļēāļ™. āļ—āļĩāļĄāļ—āļĩāđˆāļĄāļĩāļ„āļ§āļēāļĄāđ€āļ‚āđ‰āļēāđƒāļˆāļĢāđˆāļ§āļĄāļāļąāļ™āļˆāļ°āļ•āļąāļ”āļŠāļīāļ™āđƒāļˆāđ€āļāļĩāđˆāļĒāļ§āļāļąāļšāļœāļĨāļīāļ•āļ āļąāļ“āļ‘āđŒ āļ™āđ‚āļĒāļšāļēāļĒ āđāļĨāļ°āļāļēāļĢāđ€āļĢāļĩāļĒāļ™āļĢāļđāđ‰āđ„āļ”āđ‰āļ”āļĩāļ‚āļķāđ‰āļ™

āļĢāļ°āļ§āļąāļ‡

  • āđāļ•āđˆāļĨāļ°āļ—āļĩāļĄāļ­āļēāļˆāđƒāļŠāđ‰āļ„āļģāđ€āļ”āļĩāļĒāļ§āļāļąāļ™āļ•āđˆāļēāļ‡āļāļąāļ™ āļ”āļąāļ‡āļ™āļąāđ‰āļ™āļ„āļ§āļĢāļāļģāļŦāļ™āļ”āļ‚āļ­āļšāđ€āļ‚āļ•āļ•āļąāđ‰āļ‡āđāļ•āđˆāđ€āļ™āļīāđˆāļ™āđ†
  • āđ€āļāļ“āļ‘āđŒāļĄāļēāļ•āļĢāļāļēāļ™āļ­āļēāļˆāļ”āļđāđāļ‚āđ‡āļ‡āđāļāļĢāđˆāļ‡āđƒāļ™āļ‚āļ“āļ°āļ—āļĩāđˆāļ›āļĢāļ°āļŠāļīāļ—āļ˜āļīāļ āļēāļžāđƒāļ™āđ‚āļĨāļāđāļŦāđˆāļ‡āļ„āļ§āļēāļĄāđ€āļ›āđ‡āļ™āļˆāļĢāļīāļ‡āđ„āļĄāđˆāđ€āļ—āđˆāļēāļāļąāļ™
  • āļāļēāļĢāđ€āļžāļīāļāđ€āļ‰āļĒāļ•āđˆāļ­āļ„āļļāļ“āļ āļēāļžāļ‚āđ‰āļ­āļĄāļđāļĨāđāļĨāļ°āđāļœāļ™āļāļēāļĢāļ›āļĢāļ°āđ€āļĄāļīāļ™āļĄāļąāļāļŠāļĢāđ‰āļēāļ‡āļœāļĨāļĨāļąāļžāļ˜āđŒāļ—āļĩāđˆāđ€āļ›āļĢāļēāļ°āļšāļēāļ‡

āļ‚āļąāđ‰āļ™āļ•āļ­āļ™āļ•āđˆāļ­āđ„āļ›

  1. āđ€āļĢāļīāđˆāļĄāļ•āđ‰āļ™āļ”āđ‰āļ§āļĒāļ„āļģāļˆāļģāļāļąāļ”āļ„āļ§āļēāļĄāļ āļēāļĐāļēāļ˜āļĢāļĢāļĄāļ”āļēāļ‚āļ­āļ‡āļœāļĨāļĨāļąāļžāļ˜āđŒāļ—āļĩāđˆāļ„āļļāļ“āļ•āđ‰āļ­āļ‡āļāļēāļĢ
  2. āđ€āļĨāļ·āļ­āļāđ€āļĄāļ•āļĢāļīāļāļ§āļąāļ”āļ„āļ§āļēāļĄāļŠāļģāđ€āļĢāđ‡āļˆāļŦāļ™āļķāđˆāļ‡āļĢāļēāļĒāļāļēāļĢāđāļĨāļ°āđ€āļ‡āļ·āđˆāļ­āļ™āđ„āļ‚āļ„āļ§āļēāļĄāļĨāđ‰āļĄāđ€āļŦāļĨāļ§āļŦāļ™āļķāđˆāļ‡āļĢāļēāļĒāļāļēāļĢāļāđˆāļ­āļ™āļāļēāļĢāļ—āļ”āļŠāļ­āļš
  3. āļ”āļģāđ€āļ™āļīāļ™āļāļēāļĢāļ™āļģāļĢāđˆāļ­āļ‡āļ‚āļ™āļēāļ”āđ€āļĨāđ‡āļāļ”āđ‰āļ§āļĒāļ‚āđ‰āļ­āļĄāļđāļĨāļ•āļąāļ§āđāļ—āļ™ āđ„āļĄāđˆāđƒāļŠāđˆāļŠāļļāļ”āļŠāļēāļ˜āļīāļ•āļ—āļĩāđˆāļŠāļ§āļĒāļ‡āļēāļĄ
  4. āđ€āļ­āļāļŠāļēāļĢāļ—āļĩāđˆāļ‹āļķāđˆāļ‡āļāļēāļĢāļāļķāļāļ­āļšāļĢāļĄāđ€āļžāļ·āđˆāļ­āđ€āļžāļīāđˆāļĄāļ›āļĢāļ°āļŠāļīāļ—āļ˜āļīāļ āļēāļžāļāļēāļĢāļ›āļĢāļ°āļĄāļ§āļĨāļœāļĨāļ‚āļ­āļ‡ Chinchilla āļŠāđˆāļ§āļĒāđ„āļ”āđ‰ āđāļĨāļ°āļ§āļīāļ˜āļĩāļ—āļĩāđˆāļ‡āđˆāļēāļĒāļāļ§āđˆāļēāļˆāļ°āļ”āļĩāļāļ§āđˆāļē

āļŠāļģāļĢāļ§āļˆāļ•āđˆāļ­āđ„āļ›

āļˆāļ”āļŦāļĄāļēāļĒāļ‚āđˆāļēāļ§āļŸāļĢāļĩ

āļĢāļąāļšāļŠāļĢāļļāļ› AI āļĢāļēāļĒāļ§āļąāļ™

āđ€āļĢāļ·āđˆāļ­āļ‡āļĢāļēāļ§ AI āļ—āļĩāđˆāļ•āļĢāļ§āļˆāļŠāļ­āļšāđāļĨāđ‰āļ§ 3 āđ€āļĢāļ·āđˆāļ­āļ‡āļ—āļļāļāļ§āļąāļ™āļ—āļģāļ‡āļēāļ™ āđ€āļŠāđ‰āļē, āđ€āļ‚āļĩāļĒāļ™āđ€āļ›āđ‡āļ™āļ āļēāļĐāļēāļ­āļąāļ‡āļāļĪāļĐāļ˜āļĢāļĢāļĄāļ”āļē āļŸāļĢāļĩāļ•āļĨāļ­āļ”āđ„āļ› āđ„āļĄāđˆāļĄāļĩāđ‚āļ†āļĐāļ“āļē

āļ­āļĩāđ€āļĄāļĨāļŦāļ™āļķāđˆāļ‡āļ‰āļšāļąāļšāļ—āļļāļāļ§āļąāļ™āļ—āļģāļ‡āļēāļ™ āļĒāļāđ€āļĨāļīāļāļāļēāļĢāļŠāļĄāļąāļ„āļĢāđƒāļ™āļ„āļĨāļīāļāđ€āļ”āļĩāļĒāļ§ āđ€āļĢāļēāđ„āļĄāđˆāđ€āļ„āļĒāļ‚āļēāļĒāļŦāļĢāļ·āļ­āđāļŠāļĢāđŒāļ—āļĩāđˆāļ­āļĒāļđāđˆāļ­āļĩāđ€āļĄāļĨāļ‚āļ­āļ‡āļ„āļļāļ“

āļ—āļ”āļŠāļ­āļšāļ•āļąāļ§āđ€āļ­āļ‡

Take the Chinchilla Compute-Optimal Training quiz

āļ‚āđ‰āļ­āđ€āļŠāļ™āļ­āđāļ™āļ°āļ—āļąāļ™āļ—āļĩāļŠāļģāļŦāļĢāļąāļšāļ—āļļāļāļ„āļģāļ•āļ­āļš āđāļĨāļ°āđƒāļšāļĢāļąāļšāļĢāļ­āļ‡āļ—āļĩāđˆāļŠāļēāļĄāļēāļĢāļ–āđāļŠāļĢāđŒāđ„āļ”āđ‰āļžāļĢāđ‰āļ­āļĄ ID āļ—āļĩāđˆāļ•āļĢāļ§āļˆāļŠāļ­āļšāđ„āļ”āđ‰āđ€āļĄāļ·āđˆāļ­āļ„āļļāļ“āļœāđˆāļēāļ™āļŦāļĨāļąāļāļŠāļđāļ•āļĢ

āđ€āļĢāļīāđˆāļĄāđāļšāļšāļ—āļ”āļŠāļ­āļš

āļŠāļ™āļąāļšāļŠāļ™āļļāļ™āļāļēāļĢāļĻāļķāļāļĐāļē AI āļŸāļĢāļĩ AI Understanding āđ€āļ›āđ‡āļ™āļ­āļ‡āļ„āđŒāļāļĢāđ„āļĄāđˆāđāļŠāļ§āļ‡āļŦāļēāļœāļĨāļāļģāđ„āļĢ 501(c)(3) — āđ„āļĄāđˆāļĄāļĩāđ‚āļ†āļĐāļ“āļē āđ„āļĄāđˆāļĄāļĩāļāļģāđāļžāļ‡āđ€āļ‡āļīāļ™, āļ•āļĨāļ­āļ”āđ„āļ› āļšāļĢāļīāļˆāļēāļ„

āļ„āļģāļ–āļēāļĄāļ—āļĩāđˆāļžāļšāļšāđˆāļ­āļĒ

Chinchilla Compute-Optimal Training āļ„āļ·āļ­āļ­āļ°āđ„āļĢ

The Chinchilla paper estimates training-compute allocations with three complementary methods: training-curve envelopes, IsoFLOP profiles, and a parametric loss fit. It matters because its measured range and estimation methods show how researchers can cross-check a compute-optimal forecast rather than rely on one extrapolation.

How many language models did the Chinchilla study train for its scaling analysis?

The paper reports training over 400 models as part of its parameter and token sweep.

What parameter-count range did the paper’s model sweep cover?

The paper describes a sweep ranging from about 70M to over 16B parameters.

What training-token range did the paper’s sweep cover?

The study reports training runs on roughly 5B to 500B tokens.

What did the paper’s first estimation approach vary while holding model sizes fixed?

Approach 1 fixes model sizes and trains each across different horizons, then uses smoothed training curves to estimate minimum loss per FLOP.

In an IsoFLOP profile, what is held fixed while model sizes are compared?

Approach 2 compares final loss across model sizes at fixed FLOP budgets and locates the loss minimum.

āđ€āļĢāļĩāļĒāļ™āļĢāļđāđ‰āļ•āđˆāļ­āđ„āļ›

āļ„āļģāđāļ™āļ°āļ™āļģāļ—āļĩāđˆāđ€āļāļĩāđˆāļĒāļ§āļ‚āđ‰āļ­āļ‡

āļ„āļģāđāļ™āļ°āļ™āļģāđ€āļžāļīāđˆāļĄāđ€āļ•āļīāļĄāļ—āļĩāđˆāđ€āļĨāļ·āļ­āļāļŠāļģāļŦāļĢāļąāļšāļŦāļąāļ§āļ‚āđ‰āļ­āļ™āļĩāđ‰

āļ•āđˆāļ­āđ„āļ›āļ„āļģāđāļ™āļ°āļ™āļģāļ•āđˆāļ­āđ„āļ›

āļāļēāļĢāļāļķāļāļ­āļšāļĢāļĄāđ€āļ§āļĨāļēāļ—āļ”āļŠāļ­āļš

āļ„āļ§āļēāļĄāļĢāļđāđ‰āļžāļ·āđ‰āļ™āļāļēāļ™

āļĢāļąāļšāđƒāļšāļĢāļąāļšāļĢāļ­āļ‡ AI āļŸāļĢāļĩāļŦāļĨāļąāļāļŠāļđāļ•āļĢāđāļĨāļ°āļāļēāļĢāļŠāļ­āļšāļŠāļąāđ‰āļ™ āđ† āļŸāļĢāļĩ āđ„āļĄāđˆāļĄāļĩāļ„āđˆāļēāđƒāļŠāđ‰āļˆāđˆāļēāļĒ āđ„āļĄāđˆāļĄāļĩāļšāļąāļ•āļĢāđ€āļ„āļĢāļ”āļīāļ• āļ›āļĢāļ°āļĄāļēāļ“ 2.5 āļŠāļąāđˆāļ§āđ‚āļĄāļ‡āđ€āļĢāļīāđˆāļĄāļ•āđ‰āļ™āļŸāļĢāļĩ →