SLOP FIGHTER · published
Good news, I have improved narration quality significantly across the board. Besides adding a bit of movement and energy to the prepositional phrases by shifting their positions in the sentences and converting some to participial (an old trick), I have done something I probably should have done a while back and updated the dataset training methods. Claude did some restructuring through my dataset generator and had a look at the scripts that work the datasets into LoRA adapters.
Ultimately, I believe issues with clarity were coming from there. The processing method was introducing some interference (apparently, idk I am a writer of English words). The way I understand it is now the LLM hears the words better and thus can reproduce them better. It's tuned a little better. My sentences were fine.
The LoRA adapters are kind of a bottleneck, so it's important not to strangle the LLM, you know?