Feedback :)
Hi Hasta!
I've noticed you're using Nova as the base, any particular advantages in your opinion so far? Are you using grimjim's abliteration lora, instruct merge or plain Nova with the ANJIR adapter? It likes to 'half-assedly' warn or make a snide comment when zero shotting controversial content, then proceed to comply in the same message from my studying, doesn't always happen as some refusals slip through (which is fine to me). It has trouble retaining text formatting, gets confused with scenario cards: who's speaking - transition of scenes - loosely follows the requirements: Age, species, etc... when introducing new characters. The biggest weakness so far is the tendency to make characters 'villainous' or rudely OOC. The second con would be the so called 'slop' vocabulary, is this a product of the distill or dataset? I prefer XE over PJ right now. PJ talks long windily in narration and likes to reiterate itself a lot, leading to 1000+ token messages. The only other 3.1 Nova model I've used so far was a MoE, it exhibits similar OOC mean tendencies.
Thanks for reading Hasta! I look forward to the next 3.1 release!
Hi Hasta!
I've noticed you're using Nova as the base, any particular advantages in your opinion so far? Are you using grimjim's abliteration lora, instruct merge or plain Nova with the ANJIR adapter? It likes to 'half-assedly' warn or make a snide comment when zero shotting controversial content, then proceed to comply in the same message from my studying, doesn't always happen as some refusals slip through (which is fine to me). It has trouble retaining text formatting, gets confused with scenario cards: who's speaking - transition of scenes - loosely follows the requirements: Age, species, etc... when introducing new characters. The biggest weakness so far is the tendency to make characters 'villainous' or rudely OOC. The second con would be the so called 'slop' vocabulary, is this a product of the distill or dataset? I prefer XE over PJ right now. PJ talks long windily in narration and likes to reiterate itself a lot, leading to 1000+ token messages. The only other 3.1 Nova model I've used so far was a MoE, it exhibits similar OOC mean tendencies.Thanks for reading Hasta! I look forward to the next 3.1 release!
Thank you for your feedback! This model is fine-tuned directly from Supernova and without lora. The dataset used included RP data and my own 'Villainous' set, which I created to test my new training code that freezes some of the model parameters (training only 14 out of 32 layers). That’s why I used distinct data as the default persona, rather than a helpful assistant. The refusals might be occurring because I haven’t found the ideal settings yet, which is why I’m still experimenting. This approach helps me identify the best settings for future training. So, this is just one of several experiments (12 in total) to test how effective my training code is, even though it doesn’t 'fully fine-tune' the model, it influences it more than just lora. Thank you again for your feedback, and hopefully, I can release a stable version in the future