A language model capable of impeccable reasoning that consistently trips over stock phrases, stacked adjectives, and the ubiquitous ‘It’s not X, it’s Y’ – anyone working with Google’s Gemma series knows exactly what we mean. Scotoma-2, a project by Hugging Face community member AesSedai distributed in GGUF format, is a focused effort to eliminate these stylistic tics while keeping the original model’s intelligence intact.

The technical approach is noteworthy for its surgical precision. The creator first used a technique known as ‘abliteration’ – a selective suppression of unwanted patterns – to weaken what he calls the assistant persona, which he believes is the main source of the repetitiveness. Then, using J-lense projection, he worked to preserve the model’s cognitive capabilities while isolating the behaviors to be corrected. Building on this, he constructed four datasets of rejected-output vs accepted-output pairs, each targeting a specific prose issue, and performed four consecutive rounds of fine-tuning with Direct Preference Optimization (DPO). The result is a cleaned-up version that, at the same size (31 billion parameters), reads more naturally, avoids robotic phrasing, and doesn’t compromise response quality.

This isn’t just a stylistic cleanup for enthusiasts. It carries concrete implications for organizations that have adopted self-hosted inference. In an on-premise setting, the ability to sculpt the model’s tone – stripping away the mannerisms baked in during original training – means aligning it with a company’s or application’s voice without resorting to costly cloud services or accepting the rigid moderation policies of external providers. An enterprise that chooses to run the model on its own servers, perhaps for data sovereignty or to control TCO, can now apply techniques similar to Scotoma-2 to craft an assistant that not only answers accurately but does so with a coherent, recognizable voice.

There’s a broader signal here. The fact that the community is stepping up to fix model tics, using methods that go beyond standard fine-tuning and delve into internal representation manipulation, shows that expressive style control has become a key lever of differentiation. For those evaluating on-premise deployments, the capability to shape surface behavior without breaking reasoning foundations is a competitive advantage that’s hard to achieve with cloud APIs alone. It also suggests the open model ecosystem is maturing toward a sophistication that addresses real-world usability, not just abstract benchmark performance.