为什么没有针对 LLM 的 LoRA?
1 分•作者: instagraham•8 个月前
当你浏览图像生成模型的 LoRA(例如在 Civitai 上),你会发现它们提供了非常强大的视觉风格范围——这使得基础模型能够不断被优化,以适应当时的流行趋势。
那么,文本 LLM 有什么不同呢?例如,为什么没有针对 Python 编码、科幻小说写作等方面的特定 LoRA?
我读到过一些讨论,说文本 LoRA 只能针对它们的基础模型——这其实在图像生成方面也基本如此。
丰富的 LoRA 文化将使许多模型更具通用性,并减少不断下载新检查点的需求。
查看原文
When you look at LoRAs for image generation model (like on Civitai), you get a very powerful range of different visual styles - allowing a base model to be constantly refined to suit whatever is trending at the time.<p>What makes text LLMs different, then? For example, why are there no specific LoRAs for things like python coding, or sci-fi writing, or so on?<p>I've read discussions about text LoRAs being specific to their base model - which is honestly also largely the case with image gen.<p>A rich LoRA culture would make a lot of models more versatile and reduce the need to keep downloading new checkpoints