精选理由
NVIDIA 用 LoRA 微调小模型 Cosmos 3 Nano 做交通信号识别，准确率从 54% 蹿到 93%，一天内搞定，效果很猛。
零样本模型漏检了一个可见交通信号。用 LoRA 对 Cosmos 3 Nano 进行后训练后，WTS 验证准确率从 54.41% 提升到 87.14%。结合 NVIDIA TAO AutoML，准确率进一步达到 93.35%。整个微调流程在一天内完成。
原文 · NVIDIA AI
🚦 The zero-shot model misses a visible traffic signal. After post-training Cosmos 3 Nano with LoR...
🚦 The zero-shot model misses a visible traffic signal. After post-training Cosmos 3 Nano with LoRA, it correctly identifies the intersection, and overall WTS validation accuracy increases from 54.41% to 87.14%. 📈 With NVIDIA TAO AutoML, accuracy reaches 93.35%. All in under a day. 💬 9 🔄 6 ❤️ 35 👀 3251 📊 14 ⚡