Quality merge

#1
by redaihf - opened

This is a stable merge that has a sharp and distinctive thinking style. It appears to be mostly decensored although there are occasional hints of the unsafe content avoidance strategies that plague ARA-abliterated models in its CoT traces. This may explain its inability to write longer responses on request as well as its difficulties in reliably translating its insightful analyses during thinking into its final responses.

Some examples from the CoT traces:

Making excuses

3000 words (LLM constraints usually mean a condensed version)

Complaining

Ensure Lady Harrison doesn't sound purely evil—the worldbuilding requires her to be the heroine in her own twisted way

Missing material

Describe the alien cuisine in detail

Thank you for your insights! I'm sure the content avoidance was inherited from the MeroMero SFT (ApocalypseParty/G4-26B-SFT-6), which was based on the stock model, though when I put this merge, I didn't care about its uncensored aspect in the first place.

This is my first merge ever and was done in a rush, so any feedback and insight are welcome!

You could test your theory about the covert noncompliance being inherited from MeroMero by replacing the ARA base with the SOMPOA Heresy equivalent.

Sign up or log in to comment