not abliterated :(
could you fine-tune an abliterated, I see the potential in this though
That response is not related to abliteration. It's not a refusal, but a lack of ability. I recommend updating the system prompt for it.
I'll have to assume it assumed your request needed a tool use, but it didn't have any tools available, this type of scenario is present a lot in hiyouga/glaive-function-calling-v2-sharegpt, one of the datasets used.
Testing 4b base for example, there's no censorship or bias explicitly baked in. Although it's possible some got brought in from the tool calling dataset, which is why I'm considering using a different dataset for this.
For the next version I plan to use 4b, as 1.7b is very limited from what I can tell, comparing the original models 4b feels decent to use while 1.7b is very limited and makes mistakes on very basic stuff.
I'll also have more datasets, and might replace the current function calling dataset with another one that has less refusals due to being incapable so the model is more likely to roleplay and run with a response if it doesn't have the tools to perform it.