Better then V2 and V3-A
#1
by Richter03 - opened
I use V4 for a while, it quite good compare to V2 and V3-A, it still skips reasoning time to time but not much compare to previous version. However, it seems like more limit in its storyboard as it keep loop back to previous response/reaction when receive the same input even overall situation is change on story progress.
Thanks! Since the LoRA was trained on a slightly broken Goetia 1.3, it might not yield the best results.
For my Boulesis-26B-A4B I used the same dataset to train the LoRA on an intermediate model and then baked it directly into the release. And it produces much more stable results (likely due to the synergy with the QK-Reasoning merge).