How to Bake Personality Into an LLM in 15 Minutes

Logosofia no Dia a Dia · Apr 27, 2026 · 25m

Want an LLM that doesn't default to cheerful, hedged over-explaining? This episode unpacks the state-of-the-art recipe for baking real personality into model weights — not just system prompts. We break down the Grumpy Italian Chef case study (a 1.2B model trained in 15 minutes on a consumer GPU), explain the SFT + DPO pipeline, and explore how much data you actually need for style transfer vs. robust persona alignment. Plus: tooling options (Unsloth, LlamaFactory, Axolotl), the beta personality dial, and the philosophical question of whether different alignment is misalignment. Episode 2495 —…