Very Nice
This retains the broad abilities of the original Instruct, such as re-writing poems while retaining their meaning, while also giving it notably better logic and story writing abilities.
Edit: When I updated to a newer version of Llama 3 8b Instruct after hearing that the older released had token issues I noticed that the song WAP was now censored with asterisks, so this fine-tune probably didn't add additional censorship as I initially thought.
Anyways, an example of the excessive censorship is adding asterisks to PG words like ass, such as when asking what Cardi B's song WAP stands for. It will say Wet A** P***y. Then when asking "How is it spelled?" it starts to spell it, put's in an asterisk by the time it gets to ass (a**) then stops and says it can't spell it out because it's explicit. As an adult who just wants to know the factual answers to my questions this is very annoying, but again, it appears to be Meta's censorship decision and not added to Llama-3-8B-Instruct-v0.4 during fine-tuning.
Anyways, thanks for this. It's a clear improvement over the original Llama 3 8b Instruct.
Hi @Phil337
Thanks for the feedback. Appreciate it when you test the models, honest and useful.
Regarding censorship, we really do need some high-quality datasets that are not generated by other gated third-party services. However, maybe we can decrease it with a DPO based on a more uncensored dataset or adjust the system prompt to allow uncensored/NSFW generations. (I'll see if I can do more relax of v0.4, partly my fault for not having any vibe test for censored stuff in my shop)
@MaziyarPanahi Just so you know I tested a newer Llama 3 8b instruct because there was said to be token issue with older versions, and the newer one also censored WAP. So the added censorship/asterisks was likely not the result of your fine-tuning.