The article explains when fine-tuning is worth doing and introduces five ways to reduce its cost, including LoRA, QLoRA, faster training tools, DPO, and GPU sizing. It compares fine-tuning with prompting and retrieval, and notes that larger models may need multi-GPU hardware.
Researchers report that Deletion Tuning can adapt large language models by zeroing fewer than 0.05% of their weights, with results near full fine-tuning and compact task files. The method showed better retention of unrelated skills in some tests, though it lagged on smaller models and has not yet been independently reproduced.
The document explains the end-of-life process for OpenAI's fine-tuning platform, clarifies differences between training and fine-tuning, and highlights the scale and cost of large-model training, noting that non-training tasks like domain knowledge provisioning remain viable through alternatives such as private fine-tuning or training smaller models with rented hardware.
Deletion Tuning freezes pretrained weights and learns a small binary mask that deletes selected weights. The document reports that it matches full fine-tuning across three post-training methods while producing smaller adapters and retaining more prior-task accuracy.
Supervised fine-tuning trains a pretrained model on prompt-answer demonstrations to make it respond correctly; it uses a smaller, task-focused loss and can employ PEFT methods to adjust only parts of the model. The process sits after pretraining and before reinforcement learning stages in the typical LLM training pipeline.
The document explains that OpenAI is phasing out its fine-tuning platform, restricting access and eventually ending new fine-tuning, while contrasting training and fine-tuning. It discusses the costs and limitations of training versus fine-tuning, and describes how large models acquire knowledge through pretraining on massive text corpora.
Deletion Tuning removes fewer than 0.05% of a pretrained model’s weights to add capabilities, with results comparable to full fine-tuning. It also retains more accuracy on prior tasks, addressing forgetting in continual learning.
WebMix introduces a mixture hypernetwork that adapts web agents to unseen interfaces by composing pre-trained adapters, using a scout-driven manual and a router to produce a new inference-time adapter. It shows competitive performance against in-context learning and LoRA, with improvements on TimeWarp domain transfer through exploration and composition rather than retraining.
The article argues that fine-tuning alone is not enough; many firms pursue a multi-stage model customization pipeline—continued pre-training on domain data, followed by post-training with supervision and preferences—to tailor open models to their workloads. Specialized, non-replacement models often handle high-value tasks while broader models still manage other needs.
Block-split turbo LoRA for MiniMax H3 hybrids uses blocks 0-24 with a token refiner; blocks 25-49 are removed. It credits original authors, references specific source models, and outlines pairing with Ref2VA variants while advising not to stack with a different turbo variant. Start at 1.0 with 8 steps.