FlagEmbedding icon indicating copy to clipboard operation
FlagEmbedding copied to clipboard

Retrieval and Retrieval-augmented LLMs

Results 622 FlagEmbedding issues
Sort by recently updated
recently updated
newest added

作者您好,我在unified finetuning bge-m3时遇到了一些问题,希望能得到您的帮助,感谢! 1、下面是我unified微调时的参数,我微调之后发现,在我的测试集上大部分数据(大约5w条)的“稀疏向量”全变成了0,而微调前只有400条是0。请问是我的参数设置有问题吗,还是因为数据的原因呢? 2、unified finetuning后,dense embedding检索比单独的dense finetuning的效果要差一些, dense finetuning时per_device_train_batch_size=80, unified finetuning时per_device_train_batch_size=70,其他参数一致。 3、self_distill_start_step需要设置吗? 4、我有两个任务: 1)相似问题对(qq); 2)问答对(qa); qq的文本较短,集中在100以内;qa的文本较长,集中在300左右。那么unified finetuning的时候是可以直接将他们混合打乱一起训练吗,或者使用efficient batching strategy来按长度分割训练? `nohup torchrun --nproc_per_node 4 \ -m FlagEmbedding.BGE_M3.run \ --output_dir /data1/melo/embedding_models/emb_1...

When finetuning bge-large-en-v1.5: ``` Traceback (most recent call last): File "/opt/homebrew/bin/torchrun", line 8, in sys.exit(main()) ^^^^^^ File "/opt/homebrew/lib/python3.11/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 355, in wrapper return f(*args, **kwargs) ^^^^^^^^^^^^^^^^^^ File "/opt/homebrew/lib/python3.11/site-packages/torch/distributed/run.py", line 919,...

Thank you for your excellent work. I used Visualized BGE for testing on COCO/Flickr, but the retrieval performance was quite poor. However, I used the stage2 weights. I would like...

Hi, I would like to ask about incorporating additional training objectives that are beneficial to downstream tasks during the pre-training of BGE on top of the MLM task. Specifically, my...

{"4":0.055145263671875,"6":0.01322174072265625,"16":0.012908935546875,"37":0.012176513671875,xxxxxxxxxx} 相比于BM25稀疏算法得出来的分数,BGE-M3模型稀疏出来的向量分数大多在0.01以内,存入milvus内通过稀疏检索,哪怕是使用同一稀疏向量的值去检索,得出的分值也在0.02分以内,这会让我排序阈值不好控制,请问是我缺少什么参数吗?通用RAG对于稀疏向量的阈值分数是多少呢?

I am currently working on a project that involves finetuning Visualized BGE. I have been able to successfully use the pretrained model, but now I would like to further finetune...

大神们好。 我在对rerakner进行一些实验的时候,发现一些问题。(**bge-reranker-v2-m3**模型) **问题:** 1:全精度(fp32)比 半精度(fp16),整体推理耗时慢太多了(差不多5倍的差距)。请问这是正常的吗 2:我对token耗时和模型推理耗时进行了统计。发现全精度的token耗时占比太大了,比半精度大多了。正常直觉来看,不应该是全精度的模型推理耗时上升吗,为啥是token耗时(文本-->token)剧烈增加?如下图 ![1](https://github.com/user-attachments/assets/1b63ff99-86a4-4c2a-9646-f3986cbe85f9) 另外,如何才能降低 token转换 的耗时 ?

如果我的数据集每条负样本数量不一致,比如在25-35之间浮动。train_group_size应该如何设置,可以直接取20之间的值么? 还是我应该去掉一些,让每条负样本的长度一致,比如19,然后再设置train_group_size为20

A data include some pictures and a text. Like: image1,image2,image3 and a text. Can I use **visualized bge**? How to use, thank you! ``` def encode(self, image=None, text=None): # used...