Hugging Face Blog
  • Score 86
  • Official

Why your fine-tuning job takes longer than the FLOPs say

A CatBoost predictor trained on 272 production SFT runs forecasts fine-tuning wall-clock time with R²=0.98 and 1.71h MAE, far beating FLOP-based estimates (R²=0

So what

What this event means by reading role—not a longer recap.

  • BuilderIf you schedule training jobs, log config plus duration and train a similar regressor; evaluation cadence and checkpoint policy matter as much as model size.
  • ResearcherThe paper shows step-level analytical models like Calculon explain only ~11% of median wall-clock, so overhead-aware learned predictors are the better baseline.
  • InvestorBetter duration prediction directly cuts idle GPU time on scarce, expensive clusters, a concrete lever for utilization and cost in training infrastructure.

Score dimensions

Higher total means read first. Each bar is one factor we use to rank the system pool. How we score

RelevanceHow tightly this is about AI.
78
ImpactHow much this could change the field or the market.
62
NoveltyHow new this is versus a recap.
70
CredibilityHow much we trust the source.
92
ActionabilityWhether a reader can do something with it.
72