A fine-tuning approach that combines a frozen quantized model with trainable low-rank adapters.
Sign in to saveQLoRA reduces adaptation memory by storing the frozen pretrained model in a low-bit representation and training LoRA adapters through it. The paper describes four-bit quantization and techniques for managing memory during training. QLoRA is a training method, not the name of one specific downloadable model.
Research paper first submitted to arXiv on 2023-05-23.
Relevant to fine-tuning under GPU memory constraints. Actual requirements depend on model size, sequence length, batch size, and the implementation. Catalog note: this is a source-verified reference entry, not a measured current-trend signal. No search-volume or growth claim is made.
Pro opens the same-day prioritize / watch / skip table, including names that are still too new to publish here.
This name has not appeared in an imported daily report yet.