Issue 2026-11-18 · Industry · 研究 · 应用
Study: FLOP-based estimates poorly predict fine-tuning time

Using 272 production SFT jobs (~7,382 GPU-hours on NVIDIA H200), Hugging Face found the standard FLOP-based estimate explains little of observed runtime (R²=0.31, MAE 12.41h). A CatBoost regressor using only pre-dispatch information reached R²=0.98, MAE 1.71h and median error 0.51h in 5-fold CV. The result matters for scheduling shared GPU clusters.
Read original ↗- OCR Costs Fall Fivefold in 18 Months to Under $130 per Million Pages
- Chipmakers ignore frontier AI pause calls, ramp up output
- Nvidia reaffirms open-source AI strategy, Hugging Face deal on track
- XYBench benchmark tests whether AI assistants correct users' mistaken assumptions
- Huang resignation raises doubts over Nvidia's open-source commitment