A reproducible confidence interval for the compute-optimal tokens-per-parameter ratio

Chinchilla's own models imply about 20 tokens per parameter (70B parameters on 1.4T tokens). Besiroglu et al. (preprint) report that the published parametric-fit intervals were implausibly narrow, requiring over 600,000 experiments where fewer than 500 were likely run, and that a refit agrees with the other two approaches. Pearce and Song attribute Kaplan's different exponents mostly to non-embedding parameter counting and small scale. Porian et al. (arXiv:2406.19146, NeurIPS 2024) reproduced the Kaplan-versus-Chinchilla discrepancy on two datasets, released their losses and code, and give bootstrap 95% intervals for the scaling exponent of optimal model size (about 0.50); for the optimal tokens-per-parameter ratio they report only ranges across budgets (for example 14–16 on RefinedWeb with tuned hyperparameters, 11–22 on OpenWebText2), with hyperparameters tuned at 20 tokens per parameter. A 2026 community dataset pools 814 published IsoFLOP points (open-athena/isoflop-experiments) without intervals. No released sweep gives per-budget bootstrap intervals on the ratio itself under a per-run-tuned schedule.

A defensible interval tells planners how much of the compute-optimal frontier is estimation noise before they fix model size, and the released data lets others test new parametric forms.

A public dataset of final losses for every (N, D) run, code that regenerates the intervals, and a statement on whether the three approaches agree, reproduced by an independent party from the released losses.