LLLiHaotian

Results 9 issues of LLLiHaotian

是的,两个训练数据的格式是一样的,可以直接使用同一份训练数据。 建议的流程是,训练完向量模型后,用向量模型挖掘出难负样本,再去训练reranker,让reranker能更好区分向量返回的top-k结果。 _Originally posted by @staoxiao in https://github.com/FlagOpen/FlagEmbedding/issues/157#issuecomment-1754223604_

### Is there an existing issue for this? - [X] I have searched the existing issues ### Describe the bug from pymilvus import connections connections.connect("default", host="localhost", port="19530") When I pass...

kind/bug

您好,我在对BAAI/bge-m3-unsupervised进行微调过程中参照您提供的toy_finetune_data.jsonl数据格式 {"query": "Five women walk along a beach wearing flip-flops.", "pos": ["Some women with flip-flops on, are walking along the beach"], "neg": ["The 4 women are sitting on the beach.",...

请问,在用bert-base-case、chinese-bert-wwm-ext、chinese-roberta-wwm-ext、chinese-roberta-wwm-ext-large这几个预训练模型跑多标签分类实验的时候都没问题,为什么使用roberta-xlarge-wwm-chinese-cluecorpussmall这个预训练模型跑多标签分类实验,在训练过程中一直 accuracy:0.0000 micro_f1:0.0000 macro_f1:0.0000 为什么会出现这种现象?求解答

请问这种loss产生偶尔上升的情况是否正常,又该如何判断预训练合适结束? [bge-m3-patent-retromae_batch56_max350.log](https://github.com/FlagOpen/FlagEmbedding/files/15360497/bge-m3-patent-retromae_batch56_max350.log)

如果要基于自己的中英文文本数据,对BGE-M3模型进行RetroMAE的二次训练,是直接对xlm-roberta进行二次训练吗

如果想基于RetroMAE预训练bart、t5系列的模型,应该如何解决呢? [bart-base-chinese-cluecorpussmall-retromae_batch256_max350.log](https://github.com/FlagOpen/FlagEmbedding/files/15302049/bart-base-chinese-cluecorpussmall-retromae_batch256_max350.log)

您好,我注意到https://github.com/FlagOpen/FlagEmbedding/tree/master/examples/pretrain中提供了预训练示例,这里的预训练是从头开始重新预训练一个模型,不知我这样理解是否正确? 假设我理解正确的话,请问如果要针对某特殊文本(比如:专利标题文本等)这种无监督数据,对现有的预训练模型、进行基于RetroMAE算法的二次训练,这该如何实现呢? 数据格式与您提供的格式一致 {"text": "一种用于溶胶法SERS检测的微流控芯片及其使用方法"} {"text": "一种钢筋防腐用韧性涂料及其涂覆方法"}

麻烦您帮我看一下,为什么在微调bge-m3的时候会出现如此报错。 之前在没有query数量为1、pos数量为1、neg数量为10的时候为微调训练正常进行;目前调整为query数量为1、pos数量为11、neg数量为10,却有报错信息,我查看了train_data,发现没有什么问题。 微调训练命令如下 nohup \ torchrun --nproc_per_node 2 \ -m FlagEmbedding.baai_general_embedding.finetune.run \ --output_dir /bgem3/supervised_simcse_fine-tune \ --model_name_or_path /bgem3 \ --train_data query_pos_neg_data.jsonl \ --learning_rate 1e-5 \ --fp16 \ --num_train_epochs 200 \ --per_device_train_batch_size...