Unsloth fine-tunes LLMs like Qwen3.5, Gemma 4, and Llama 3.2 into decision models that score provided options and return a calibrated probability choice instead of generating text.

LoRA training (r=16, one epoch) raises downstream accuracy from 30–37% to 77–81% across models of 2.5–14.4GB in 30–50 minutes. Training runs via the Unsloth desktop app, a Python API using FastDecisionModel and DecisionTrainer, or free Google Colab and Kaggle notebooks.