Fine-tuned vs from-scratch training times on two unseen distributions.
Fine-tuning reaches within 1% of from-scratch performance 20× sooner on A
and 4× sooner on B.
Figure 4. Dashed lines mark 1% below each from-scratch result; the labeled
fine-tune point is where the curve crosses it. Note the logarithmic x-axis;
training time is in 64×A6000-equivalent hours. Hover or focus the chart and
use arrow keys for exact values; the table view lists them all.