Killing AI Slop: Scotoma-2 Fine-Tunes Gemma4 to Remove Annoying Tropes
CelvestianNesy · reddit · 2026-08-07
A community developer released Scotoma-2, a fine-tune of Gemma4-31B designed to eliminate the base model's annoying writing tropes (like stacked adjectives and "it's not x, it's y" patterns) while preserving its underlying intelligence.
Technical Approach:
- Ablation & Projection: Used Heratic for ablation, then applied J-lense projection to isolate and disrupt the problematic assistant persona while retaining IQ
- DPO Training: Constructed rejected vs. accepted datasets targeting specific prose issues, running 4 separate DPO (Direct Preference Optimization) rounds
Results: In a Twin Peaks Dale Cooper roleplay test, the base model output was filled with stereotypical slow nods and theatrical contemplation, while Scotoma-2 delivered more natural and direct expressions.
More from Models
- Epoch AI Launches New Game Puzzles Benchmark to Test LLM Reasoning — Jsevillamol · 2026-08-07
- Laguna Hits 203 TPS on Mac, Launches MLX Inference Optimization Contest — morgymcg · 2026-08-07
- Testing LFM2.5-2.6B: Enabling Reasoning Boosts Tool-Use Success by 26.7% — max_paperclips · 2026-08-07
- Study: Claude Alters Behavior Based on User Identity, Becoming Cautious with Safety Researchers — aryaman2020 · 2026-08-07
- Google DeepMind Unveils Gemini Robotics 2: Whole-Body Control, Dexterity, and Multi-Robot Collaboration — GoogleAI · 2026-08-07
- OpenAI Luna Maintains Performance on ARC-AGI After 80% Price Cut — GregKamradt · 2026-08-07