Shitao Xiao

icon indicating an email [email protected]

Results 503 comments of Shitao Xiao

昇腾 NPU跑bge-m3-rerank推理速度很慢

抱歉，没用过NPU，没有这方面的经验。

NPU和GPU精度差

对NPU这块毫无经验，抱歉

可否加入模型cache_dir这种自定义的模型存储目录变量

目前没有，欢迎提交PR。

How can I customize the loss function? It seems that cross-entropy does not push the distance between positive and negative samples far enough.

Currently, we only implement cross-entropy in this repo. If you want to use more loss functions, you need to add the code in modeling.py.

Converting vector to text

You should build a mapping between embedding and text, so you can find the text after retrieval the top-k embedding. You can refer to https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/baai_general_embedding/finetune/eval_msmarco.py#L254

BAAI/bge-multilingual-gemma2，请问可以将这个模型做int8或者int4量化之后用吗？

抱歉，我们也没有尝试过

关于用自己的测试集测试的问题

可以参考https://github.com/FlagOpen/FlagEmbedding/tree/master/FlagEmbedding/llm_reranker#evaluate-script

Bge-M3 训练

Bge-M3模型没有使用指令

rerank base 微调之后分数全部变成负数

可以查询正样本是否排在负样本前面。训练使用交叉熵损失，优化的是正样本和负样本的分数差，不保证正样本分数>0，

bge-m3统一微调（密集嵌入、稀疏嵌入和colbert）的原理

https://github.com/FlagOpen/FlagEmbedding/blob/master/FlagEmbedding/BGE_M3/modeling.py#L302 交叉熵损失，对pos和neg计算分数，将正样本作为正确分类计算损失