TQ4_1S on model weights with minimal quality loss is really great. The MR discussion thread with results is specially interesting, with some models much more impacted than others in PPL increase, possibly size and model architecture play a part. Are there consolidated learnings from all the experiments? Thanks for this!