• Sources: Unsloth documentation, HN discussion
  • Summary: Unsloth documents Dynamic 3.0 GGUF quantization as a pure post-training method, with no fine-tuning or retraining step. The calibration importance matrix is published for inspection rather than described. Accuracy is reported through a divergence metric scored on prompts held out of the calibration set.
  • Why it matters: The quantization is pure post-training, the calibration data is published rather than described, and the new metric is scored on prompts held out of calibration, so the accuracy claim is checkable rather than asserted.