Do you think this could be adapted to optimize the Copier training method? It can train characters and concepts on very little data pretty quickly and precisely as long as there is difference between two or more examples. Also has the occasional benefit of creating "slider" LoRA this way. The only issue is filesizes are still quite large as is typical. Also with how fast ia3 goes I wonder if it could make this process go even faster.

https://rentry.org/d2oqv

https://note.com/emanon_14/n/n7eeb65472eb5

Also any reason for a rentry instead of an article? It might give it better visibility over here.

Am i retarded or is your guide not working. If i get the model and use it, the input is the same as the standard generation

Now listen here Prodigy guy (you twisted firestarter), I don't doubt that you've gotten great results. I don't doubt your assertion that iA3+Prodigy is the best LoRA-thing since sliced cheese. However, I don't think your settings, optimized though they may be, are nearly as flexible across datasets as you think they are. I've been following your rentry for a while now and have tried many iterations of your settings, but none have produced usable results with my dataset compared to standard LoRA settings with AdamW, Adafactor, or Lion. I've tried all your suggestions for adjusting settings to accommodate the dataset as well as edited the dataset to your recommendations, but still, I've gotten no usable results with your settings or any variation thereof.

Is there anywhere you post to discuss your experiments and findings? Any Discords you participate in? I know at least one Discord filled with like-minded people who would be absolutely fascinated to hear from you and put your methods into practice. Nothing like the disbelievers on whom you spit hot fire in your rentry.