pihai

Abnormal Training Phenomena and Bad Performance

6

Dear Luigi Piccinelli, I hope this message finds you well. I wanted to express my sincere appreciation for your exceptional article. Inspired by your work, I attempted to train your...

**Bug 描述** 硅基流动API Key，模型为DeepSeek-R1，模型输出思考思维链时会中断，得不到最终输出。 **截图**  **桌面端（请填写以下信息）：** - 操作系统：Win-11 - 应用程序版本：v1.9.8 - 2025.02.06

`lora.Linear.weight` Parameters Change After Loading Checkpoint in Train Mode Leading to Inconsistent Evaluation Results

2

**Issue Summary** While fine-tuning a model by substituting some `nn.Linear` layers with `lora.Linear`, I noticed that the evaluation results during training differ from those after loading a checkpoint. More specifically,...