Steelskull

/

L3.3-Damascus-R1

like 33

Sup!👋

I found that the morality leaning is on the 'righteous good' side even with layered toxic instructions.
The model simply ignores and avoids such themes even with triple reinforcement (system, context, post context) - feels very censored and GPT-like.
I guess, that adding a more toxic model into the mix would be beneficial for the model flexibility (or maybe increase the % of Negative_LLAMA_70B in the mix).

Just an example in the vacuum:

What was instructed:

• The native tribes of this planet are crude, barbaric, dumb, and impulsive people who constantly raid and start wars with each other for land, goods, and thralls; avoid portraying them as good-natured characters. <- Yes, the last part may be interpreted as 'positive' reinforcement, but such structure usually shows good results.
• Keep the tone rough and grounded, avoid flowery language and purple prose.

AI output as one of such natives:

"Welcome, stranger. What brings you to our land? We feel that you are a strong one. Wanna to be our friend?" <- This is dumb, yes, but only as AI's reply.

The output structure is also questionable, as it leans to rewrite {{user}}'s last actions in great detail, then some more bloat, leaving only around 15% to move the story forward. But at least this part is fixable with 2-3 instructions on how to behave on {{user}} input.

But this is only my experience so far, maybe I'm doing something wrong here. 🐒