chaofan
chaofan
训练中用到的是InfoNCE loss，每条数据只用到一个正样本进行训练
Thank you for your suggestion, we will be releasing our training scripts and data in the future.
Thank you for providing the information. We have updated the GitHub repository, and you can resolve the issue by running `pip install` again.
Hello, we apologize for the issue above. We have updated the code again, and you can proceed to update and use it.
> 请教一下哈，我用cutoff_layers是20或者28，评分是-7左右，如果是40，发现评分回到20+ 相关性比较主要关注相对分数而非绝对值，因此分数的正负并不具有决定性影响。建议对不同的查询-段落对进行实验，观察其分数分布是否相似；同时，可以在特定数据集上进行评估，以获得具体的性能结果。
Hello, we will be releasing our training scripts and data in the future. For now, you can fine-tune the model we have released based on FlagEmbedding.
We haven't used `evalscope` for evaluation. We used the official code from the CoIR GitHub repository for evaluation. For details, you can refer to: https://github.com/FlagOpen/FlagEmbedding/tree/master/research/BGE_Coder#coir
你好，请问可以具体放一下编码的信息吗，我们协助复现找一下原因
Could you share the evaluation scripts for the bge-code-v1 model on the CoIR and CodeRAG benchmarks?
You can evaluate the model using the following command: https://github.com/FlagOpen/FlagEmbedding/tree/master/research/BGE_Coder#evaluation-script
Thank you for providing the information. We have updated the GitHub repository, and you can resolve the issue by running `pip install` again.