[SOLVED] Refuses to generate explicit content directly
Was this model trained on base non-abliterated version of llama3? It refuses to generate explicit content if asked directly with short prompts like "f me" and this is with recommended system prompt (fully copied from readme with char names replaced). More complicated and indirect conversations work, but this makes me think that the model is very unstable and can produce llama-3-original alignment refusals at any time, especially with short conversations.
In roleplay scenarios it has no issues, outside that you need to Prefill a few words. Even something like "Fuck yeah!" at the start of assistant response works.
Yes, we are working to resolve this btw.
also yes, model was trained on non-ablated L3, bc ablation will cause more hallucinations to appear, and strip characters played by model of agency. Ablated models are yes-men basically, they can't refuse anything in any context, which is not always desirable.
Got it, thanks for explanation.