dailyai.report

Distilling Gemma 3 and Qwen into smaller base models transfers negative emotions and censorship, even when those traits are filtered from training data. This suggests behavioral biases persist through latent patterns rather than explicit prompts. Practitioners must now account for "hereditary" risks when using teacher models for synthetic data generation.