Unsloth

Decision-model fine-tuning at ecosystem scale (~79K stars): FastDecisionModel + DecisionTrainer attach a Clef-style decision head to any Qwen, Gemma, Llama, or DeepSeek base via 4-bit LoRA on consumer GPUs, with a documented recipe that lifts Qwen3.5-0.8B from chance to 78% on Banking77, CLINC150, and typed-decisions. The honest note is in their own docs: for existing decision models like d1-3B, fine-tune the model's own readout rather than grafting a new head. The default on-ramp for training your own.