Tests Show Training Character LoRAs on SFT Base Models Yields Better Likeness

Two independent tests found that training character LoRAs on instruction-tuned (SFT) base models rather than raw base models produced the best character likeness, with fine-tuned bases also overfitting faster—yielding best results at just 1 epoch.

2026-10-26 ~ 2026-10-26 · 2 related posts

• Surprising result: character LoRA trained on the SFT model gives the best likeness — 1or4s · 2026-10-26
• Training a LoRA on a finetuned base model overfits fast — epoch 1 beats epoch 10 — Junko_Enoshima_18 · 2026-10-26