Train your own Decision Model with Unsloth (unsloth.ai)

Unsloth has introduced a groundbreaking feature that allows users to train their own decision models using fine-tuned versions of popular large language models (LLMs) like Qwen3.8 and Gemma 4. This capability transforms these models from mere text generators into decision-making systems that assess inputs, evaluate options, and deliver a selected choice supported by a probability score. The significant enhancement revealed through benchmarks shows substantial accuracy improvements, with models like Qwen3.5-2B advancing from an initial accuracy around 33% to as high as 81% after fine-tuning, showcasing the potential of this method for decision-making tasks. The new training process employs Low-Rank Adaptation (LoRA) techniques for efficient fine-tuning and can be conducted entirely for free using resources like Google Colab or Kaggle. This accessibility opens the door for more developers and researchers in the AI/ML community to explore decision-oriented AI, enhancing the robustness of their applications and potentially expanding use cases in industries ranging from finance to healthcare. Users can tap into a guided workflow to prepare data, initiate training, and deploy their own decision models with intuitive interfaces, further democratizing advanced AI capabilities.