Scotoma-2 fine-tune strips Gemma4's AI slop, keeps its brains

Scotoma-2, a community fine-tune of the open Gemma4-31B-IT model, tackles one of the most common complaints about modern LLMs: the mechanical 'AI slop' that creeps into generated prose. Created by Hugging Face user AesSedai, the model targets specific recurring sentence patterns that plague the base Gemma4, such as the overused 'It's not x, it's y' construction and heavy adjective stacking like 'a slow knowing smirk.' The goal was to strip away these stylistic tells without degrading the model's underlying intelligence.

Technically, AesSedai combined two complementary approaches. First, they used Heratic to abliterate the model—selectively removing certain learned behaviors—and then applied J-lense projection to preserve the model's reasoning capacity while isolating and disrupting the default 'assistant persona' that Aes believes drives those tics. Unsatisfied with the result from the first version, Scotoma V1, they pushed further by building four separate datasets of rejected vs. accepted outputs, each targeting a distinct prose problem. These were used for four rounds of DPO (direct preference optimization) fine-tuning. The result, as shown in sample outputs, is prose that retains the same intelligence and detail but reads far more naturally—swapping 'He doesn't look skeptical; rather, he looks…' for a firmer, punchier 'He pauses, taking a thoughtful bite of his cherry pie.' The model is available as GGUF files on Hugging Face, making it easy to run locally via llama.cpp and similar tools.