Optimal supervised fine-tuning loss depends on model capability

Original: Check out our poster session tomorrow!

Source: x.com ↗