Shitao Xiao
Shitao Xiao
huggingface的tainer应该会自动开启wandb。加上这些代码是否会造成功能的冲突？
您好，该数据来自DuReader的论文的实验章节，他使用了covid数据做检索。
任务太简单了，需要提高任务难度。没有使用neg，train_group_size应该>1（指的是pos+neg的数量，pos数量一直为1，因此train_group_size-1为从neg中采样数量）。另外可以提高per_device_train_batch_size，这也能加大任务难度。 不应该提高temperture，推荐0.01-0.1。temperture过高会让模型难以优化。我们的经验中，0.02就很好了。
并不觉得不正常。loss只是0-0.15之间波动，波动范围很小。
Don't install both cpu and gpu versions for faiss. You can uninstall faiss-cpu and re-install faiss-gpu.
You can try to upgrade the datasets package.
Currently, this script doesn't support encoder-decoder architecture.
We have no plan for this, but welcome PR from the community.
I think this error is caused by faiss. You can use faiss on CPU instead of GPU
The data format is: ``` {"content": "A is ..."} {"content": "B is ..."} {"content": "C is ..."} {"content": "Panda is ..."} {"content": "... is A"} ``` , where each line...