开源模型微调后在特定任务上超越前沿模型,成本低约 20 倍
有人发现用 DeepSeek、Gemma、Qwen 微调的小模型跑特定任务比前沿模型还强,成本只有 1/20,做垂直应用可以参考。
Bindu Reddy 发帖称,针对具体用例的定制微调模型在表现上已超过前沿模型。这些微调模型需要定期维护和更新,但成本约为前沿模型的 1/20。他提到 DeepSeek、Gemma 和 Qwen 是很好的基座模型。
Open-source fine-tunes outperform frontier models
Custom fine-tunes for specific use cases are outperforming frontier models
These fine-tunes have to be maintained and updated regularly, but are about 20x cheaper than frontier
DeepSeek, Gemma and Qwen are great based models