readme.md

用于演示多语言检索的demo

下载

在一个bash脚本里运行以下代码

git clone https://hf-mirror.com/datasets/cfli/ret_demo
# git clone 的数据不完整，需要全部删除，然后重新下载
for lang in "ar" "bn" "de" "en" "es" "fa" "fi" "fr" "hi" "id" "ja" "ko" "ru" "sw" "te" "th" "yo" "zh"
do
    # 删除旧文件
    rm -f ./ret_demo/data/${lang}/corpus.jsonl
    rm -f ./ret_demo/data/${lang}/dev_qrels.jsonl
    rm -f ./ret_demo/data/${lang}/dev_queries.jsonl
    rm -f ./ret_demo/emb/${lang}/corpus.npy
done

for lang in "ar" "bn" "de" "en" "es" "fa" "fi" "fr" "hi" "id" "ja" "ko" "ru" "sw" "te" "th" "yo" "zh"
do
    # 下载并移动文件
    wget https://hf-mirror.com/datasets/cfli/ret_demo/resolve/main/emb/${lang}/corpus.npy
    mv corpus.npy ./ret_demo/emb/${lang}/

    wget https://hf-mirror.com/datasets/cfli/ret_demo/resolve/main/data/${lang}/corpus.jsonl
    mv corpus.jsonl ./ret_demo/data/${lang}/

    wget https://hf-mirror.com/datasets/cfli/ret_demo/resolve/main/data/${lang}/dev_qrels.jsonl
    mv dev_qrels.jsonl ./ret_demo/data/${lang}/
    
    wget https://hf-mirror.com/datasets/cfli/ret_demo/resolve/main/data/${lang}/dev_queries.jsonl
    mv dev_queries.jsonl ./ret_demo/data/${lang}/
done

环境依赖

pip install gradio
pip install -U FlagEmbedding
pip install https://github.com/kyamagu/faiss-wheels/releases/download/v1.7.3/faiss_gpu-1.7.3-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

使用

1. 如果模型在之前未下载，运行app.py之前先设置export HF_ENDPOINT=https://hf-mirror.com

2. 如果要先下载模型到本地某个指定路径，可以按如下代码下载

   import os
   os.environ['HF_ENDPOINT'] = 'https://hf-mirror.com'
   
   save_path = './save_model'
   
   from transformers import AutoTokenizer, AutoModel
   tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-multilingual-gemma2')
   model = AutoModel.from_pretrained('BAAI/bge-multilingual-gemma2')
   
   tokenizer.save_pretrained(save_path)
   model.save_pretrained(save_path)
   

3. 需要修改utils.py文件第 23 行模型名称代码，将self_model_path改为【上述save_path 或 self_model_path=model_path】

4. 如果要用CPU加载模型，需要修改utils.py文件第 25-45 行

   def load_model_util(previous_model, model_path):
       self_model_path = '' # 使用上述 save_path，或者 self_model_path = model_path
       if model_path == 'BAAI/bge-multilingual-gemma2':
           if previous_model is not None and previous_model.model_name_or_path == self_model_path:
               return previous_model
           model = FlagLLMModel(self_model_path,
                               query_instruction_for_retrieval="Given a question, retrieve Wikipedia passages that answer the question.",
                               query_instruction_format="<instruct>{}\n<query>{}",
                               use_fp16=False,
                               devices=['cpu'])
       else:
           if previous_model is not None and previous_model.model_name_or_path == model_path:
               return previous_model
           model = FlagAutoModel.from_finetuned(model_path,
                                                use_fp16=False,
                                                devices=['cpu'])
       if previous_model is not None:
           del previous_model
       return model
   

5. 需要修改app.py第 11 行和第 12 行data_dir和index_dir的值，指向本地的数据/索引路径

6. 如果需要保存每次的faiss index，修改app.py第 81 行，设置faiss.write_index(faiss_index, index_path)（读index与构造index时间相近，保存不是很必要）

7. 运行app.py