Frozen Weights
الأوزان المُجمَّدة
أوزان النموذج التي تُثبَّت أثناء التدريب فلا تتلقى أي تحديثات من متجهات الميل — في QLoRA يُجمَّد النموذج الأساسي بالكامل بصيغة NF4.
Model weights that are fixed during training and receive no gradient updates — in QLoRA the entire base model is frozen in NF4 format.
Also translated asالأوزان الثابتة، المعاملات غير القابلة للتحديث، المعاملات غير القابلة للتدريب
First appears in this corpus in: Parameter-Efficient Transfer Learning for NLP (2019)
Appears in these papers
- Parameter-Efficient Transfer Learning for NLP2019in the sky ✦
- Parameter-Efficient Transfer Learning for NLP2019in the sky ✦
- BLIP-2: Bootstrapping Language-Image Pre-Training with Frozen Image Encoders and Large Language Models2023in the sky ✦
- Flamingo: a Visual Language Model for Few-Shot Learning2022in the sky ✦
- LoRA: Low-Rank Adaptation of Large Language Models2021in the sky ✦
- The Power of Scale for Parameter-Efficient Prompt Tuning2021in the sky ✦
- QLoRA: Efficient Finetuning of Quantized LLMs2023in the sky ✦
- RETRO: Improving Language Models by Retrieving from Trillions of Tokens2022in the sky ✦