Dharma-AI (Hugging Face blog) · Fine-tuning Time Predictor

Predicting End-to-End Fine-Tuning Time for Generative Language Models

Dharma-AI present a preview of a paper on predicting wall-clock time for supervised fine-tuning jobs, using 272 production runs on shared NVIDIA H200 GPUs (about 7,382 GPU-hours). A CatBoost regressor that uses only pre-dispatch configuration features reaches R² 0.98 and MAE 1.71 h in 5-fold cross-validation, compared with R² 0.31 and MAE 12.41 h for a FLOP-based baseline at 30% MFU. The full arXiv version is described as forthcoming.

Open the source ↗

Similar Research & Safety releases

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.