What Fine-Tuning an 8B Model on 250 Security Examples Actually Taught It

Research indicates that fine-tuning an 8B model on fewer than 250 security examples using consumer hardware can negatively impact performance in specific ways. This case study highlights known phenomena, specifically catastrophic forgetting, where training on a narrow safety-classification task causes the model to lose previous knowledge. While not a new discovery, this documented example provides granularity on how small-scale fine-tuning affects model stability, an active area of research within the current AI literature.