--- tags: - mteb model-index: - name: Zhihui_LLM_Embedding results: - task: type: Retrieval dataset: type: C-MTEB/CmedqaRetrieval name: MTEB CmedqaRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 29.012 - type: map_at_10 value: 41.998000000000005 - type: map_at_100 value: 43.821 - type: map_at_1000 value: 43.924 - type: map_at_3 value: 37.804 - type: map_at_5 value: 40.025 - type: mrr_at_1 value: 43.536 - type: mrr_at_10 value: 51.413 - type: mrr_at_100 value: 52.329 - type: mrr_at_1000 value: 52.366 - type: mrr_at_3 value: 49.058 - type: mrr_at_5 value: 50.291 - type: ndcg_at_1 value: 43.536 - type: ndcg_at_10 value: 48.693 - type: ndcg_at_100 value: 55.644000000000005 - type: ndcg_at_1000 value: 57.354000000000006 - type: ndcg_at_3 value: 43.627 - type: ndcg_at_5 value: 45.462 - type: precision_at_1 value: 43.536 - type: precision_at_10 value: 10.552999999999999 - type: precision_at_100 value: 1.624 - type: precision_at_1000 value: 0.184 - type: precision_at_3 value: 24.314 - type: precision_at_5 value: 17.299 - type: recall_at_1 value: 29.012 - type: recall_at_10 value: 59.123000000000005 - type: recall_at_100 value: 87.783 - type: recall_at_1000 value: 99.078 - type: recall_at_3 value: 43.474000000000004 - type: recall_at_5 value: 49.557 - task: type: Retrieval dataset: type: C-MTEB/CovidRetrieval name: MTEB CovidRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 73.235 - type: map_at_10 value: 80.87100000000001 - type: map_at_100 value: 81.10300000000001 - type: map_at_1000 value: 81.105 - type: map_at_3 value: 79.171 - type: map_at_5 value: 80.163 - type: mrr_at_1 value: 73.235 - type: mrr_at_10 value: 80.80000000000001 - type: mrr_at_100 value: 81.024 - type: mrr_at_1000 value: 81.026 - type: mrr_at_3 value: 79.15299999999999 - type: mrr_at_5 value: 80.133 - type: ndcg_at_1 value: 73.34 - type: ndcg_at_10 value: 84.387 - type: ndcg_at_100 value: 85.348 - type: ndcg_at_1000 value: 85.411 - type: ndcg_at_3 value: 80.97 - type: ndcg_at_5 value: 82.757 - type: precision_at_1 value: 73.34 - type: precision_at_10 value: 9.631 - type: precision_at_100 value: 1.005 - type: precision_at_1000 value: 0.101 - type: precision_at_3 value: 28.837000000000003 - type: precision_at_5 value: 18.209 - type: recall_at_1 value: 73.235 - type: recall_at_10 value: 95.311 - type: recall_at_100 value: 99.473 - type: recall_at_1000 value: 100 - type: recall_at_3 value: 86.091 - type: recall_at_5 value: 90.411 - task: type: Retrieval dataset: type: C-MTEB/DuRetrieval name: MTEB DuRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 28.059 - type: map_at_10 value: 85.309 - type: map_at_100 value: 87.926 - type: map_at_1000 value: 87.945 - type: map_at_3 value: 59.862 - type: map_at_5 value: 75.345 - type: mrr_at_1 value: 93.30000000000001 - type: mrr_at_10 value: 95.624 - type: mrr_at_100 value: 95.647 - type: mrr_at_1000 value: 95.649 - type: mrr_at_3 value: 95.42500000000001 - type: mrr_at_5 value: 95.572 - type: ndcg_at_1 value: 93.30000000000001 - type: ndcg_at_10 value: 91.338 - type: ndcg_at_100 value: 93.38 - type: ndcg_at_1000 value: 93.57 - type: ndcg_at_3 value: 90.512 - type: ndcg_at_5 value: 89.617 - type: precision_at_1 value: 93.30000000000001 - type: precision_at_10 value: 43.169999999999995 - type: precision_at_100 value: 4.868 - type: precision_at_1000 value: 0.49100000000000005 - type: precision_at_3 value: 80.7 - type: precision_at_5 value: 68.12 - type: recall_at_1 value: 28.059 - type: recall_at_10 value: 91.949 - type: recall_at_100 value: 98.777 - type: recall_at_1000 value: 99.816 - type: recall_at_3 value: 61.699000000000005 - type: recall_at_5 value: 79.134 - task: type: Retrieval dataset: type: C-MTEB/EcomRetrieval name: MTEB EcomRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 57.599999999999994 - type: map_at_10 value: 67.169 - type: map_at_100 value: 67.654 - type: map_at_1000 value: 67.663 - type: map_at_3 value: 64.833 - type: map_at_5 value: 66.298 - type: mrr_at_1 value: 57.599999999999994 - type: mrr_at_10 value: 67.169 - type: mrr_at_100 value: 67.654 - type: mrr_at_1000 value: 67.663 - type: mrr_at_3 value: 64.833 - type: mrr_at_5 value: 66.298 - type: ndcg_at_1 value: 57.599999999999994 - type: ndcg_at_10 value: 71.95899999999999 - type: ndcg_at_100 value: 74.092 - type: ndcg_at_1000 value: 74.323 - type: ndcg_at_3 value: 67.212 - type: ndcg_at_5 value: 69.892 - type: precision_at_1 value: 57.599999999999994 - type: precision_at_10 value: 8.7 - type: precision_at_100 value: 0.9650000000000001 - type: precision_at_1000 value: 0.098 - type: precision_at_3 value: 24.7 - type: precision_at_5 value: 16.14 - type: recall_at_1 value: 57.599999999999994 - type: recall_at_10 value: 87 - type: recall_at_100 value: 96.5 - type: recall_at_1000 value: 98.3 - type: recall_at_3 value: 74.1 - type: recall_at_5 value: 80.7 - task: type: Retrieval dataset: type: C-MTEB/MMarcoRetrieval name: MTEB MMarcoRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 72.858 - type: map_at_10 value: 81.559 - type: map_at_100 value: 81.808 - type: map_at_1000 value: 81.813 - type: map_at_3 value: 80.018 - type: map_at_5 value: 81.04299999999999 - type: mrr_at_1 value: 75.27199999999999 - type: mrr_at_10 value: 81.989 - type: mrr_at_100 value: 82.202 - type: mrr_at_1000 value: 82.206 - type: mrr_at_3 value: 80.647 - type: mrr_at_5 value: 81.53399999999999 - type: ndcg_at_1 value: 75.27199999999999 - type: ndcg_at_10 value: 84.772 - type: ndcg_at_100 value: 85.79599999999999 - type: ndcg_at_1000 value: 85.925 - type: ndcg_at_3 value: 81.884 - type: ndcg_at_5 value: 83.60300000000001 - type: precision_at_1 value: 75.27199999999999 - type: precision_at_10 value: 10.017 - type: precision_at_100 value: 1.051 - type: precision_at_1000 value: 0.106 - type: precision_at_3 value: 30.578 - type: precision_at_5 value: 19.261 - type: recall_at_1 value: 72.858 - type: recall_at_10 value: 94.197 - type: recall_at_100 value: 98.634 - type: recall_at_1000 value: 99.63499999999999 - type: recall_at_3 value: 86.6 - type: recall_at_5 value: 90.692 - task: type: Retrieval dataset: type: C-MTEB/MedicalRetrieval name: MTEB MedicalRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 56.2 - type: map_at_10 value: 62.198 - type: map_at_100 value: 62.794000000000004 - type: map_at_1000 value: 62.829 - type: map_at_3 value: 60.699999999999996 - type: map_at_5 value: 61.660000000000004 - type: mrr_at_1 value: 56.49999999999999 - type: mrr_at_10 value: 62.348000000000006 - type: mrr_at_100 value: 62.944 - type: mrr_at_1000 value: 62.979 - type: mrr_at_3 value: 60.85 - type: mrr_at_5 value: 61.809999999999995 - type: ndcg_at_1 value: 56.2 - type: ndcg_at_10 value: 65.19200000000001 - type: ndcg_at_100 value: 68.341 - type: ndcg_at_1000 value: 69.392 - type: ndcg_at_3 value: 62.163999999999994 - type: ndcg_at_5 value: 63.894 - type: precision_at_1 value: 56.2 - type: precision_at_10 value: 7.46 - type: precision_at_100 value: 0.899 - type: precision_at_1000 value: 0.098 - type: precision_at_3 value: 22.133 - type: precision_at_5 value: 14.12 - type: recall_at_1 value: 56.2 - type: recall_at_10 value: 74.6 - type: recall_at_100 value: 89.9 - type: recall_at_1000 value: 98.4 - type: recall_at_3 value: 66.4 - type: recall_at_5 value: 70.6 - task: type: Retrieval dataset: type: C-MTEB/T2Retrieval name: MTEB T2Retrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 28.79 - type: map_at_10 value: 81.687 - type: map_at_100 value: 85.19200000000001 - type: map_at_1000 value: 85.232 - type: map_at_3 value: 57.145999999999994 - type: map_at_5 value: 70.491 - type: mrr_at_1 value: 92.21000000000001 - type: mrr_at_10 value: 94.303 - type: mrr_at_100 value: 94.368 - type: mrr_at_1000 value: 94.37 - type: mrr_at_3 value: 93.94500000000001 - type: mrr_at_5 value: 94.175 - type: ndcg_at_1 value: 92.21000000000001 - type: ndcg_at_10 value: 88.29599999999999 - type: ndcg_at_100 value: 91.268 - type: ndcg_at_1000 value: 91.645 - type: ndcg_at_3 value: 89.031 - type: ndcg_at_5 value: 88.075 - type: precision_at_1 value: 92.21000000000001 - type: precision_at_10 value: 43.775 - type: precision_at_100 value: 5.097 - type: precision_at_1000 value: 0.518 - type: precision_at_3 value: 77.708 - type: precision_at_5 value: 65.473 - type: recall_at_1 value: 28.79 - type: recall_at_10 value: 87.457 - type: recall_at_100 value: 97.21499999999999 - type: recall_at_1000 value: 99.14 - type: recall_at_3 value: 58.606 - type: recall_at_5 value: 73.52300000000001 - task: type: Retrieval dataset: type: C-MTEB/VideoRetrieval name: MTEB VideoRetrieval config: default split: dev revision: None metrics: - type: map_at_1 value: 67 - type: map_at_10 value: 75.44999999999999 - type: map_at_100 value: 75.771 - type: map_at_1000 value: 75.776 - type: map_at_3 value: 73.867 - type: map_at_5 value: 74.837 - type: mrr_at_1 value: 67 - type: mrr_at_10 value: 75.44999999999999 - type: mrr_at_100 value: 75.771 - type: mrr_at_1000 value: 75.776 - type: mrr_at_3 value: 73.867 - type: mrr_at_5 value: 74.837 - type: ndcg_at_1 value: 67 - type: ndcg_at_10 value: 79.313 - type: ndcg_at_100 value: 80.894 - type: ndcg_at_1000 value: 80.989 - type: ndcg_at_3 value: 76.08500000000001 - type: ndcg_at_5 value: 77.845 - type: precision_at_1 value: 67 - type: precision_at_10 value: 9.13 - type: precision_at_100 value: 0.987 - type: precision_at_1000 value: 0.099 - type: precision_at_3 value: 27.500000000000004 - type: precision_at_5 value: 17.36 - type: recall_at_1 value: 67 - type: recall_at_10 value: 91.3 - type: recall_at_100 value: 98.7 - type: recall_at_1000 value: 99.4 - type: recall_at_3 value: 82.5 - type: recall_at_5 value: 86.8 license: cc-by-nc-4.0 language: - zh library_name: transformers --- ## Zhihui_LLM_Embedding ### Model Introduction **Zhihui_LLM_Embedding** is an embedding model specifically designed to enhance Chinese text retrieval capabilities. It is built on a 7B LLM and enhanced bidirectional attention mechanism to improved contextual understanding. The model is trained on an extensive corpus from various fields within an extremely large batch. **Zhihui_LLM_Embedding** excels in retrieval tasks, ranking **1st position** on the C-MTEB leaderboard with a leading performance score of **76.74** as of June 25, 2024. ### Optimization points * Data source enhancement: Leverages the knowledge of LLMs through three types of distillation methods.(GPT3.5 & GPT4) * Data Refinement: LLM scores candidate positive passages to select the most relevant examples. * Query Rewriting: LLM generates queries that can be answered by positive documents but are unrelated to negatives, thus enhancing the query's quality and diversity. * Query Expansion: Queries are expanded based on multiple topics for long documents. * Negative example mining: Use multiple methods and different ranges of negative selection to mine hard negative examples. * Improved Contrastive Loss: Design a novel InfoNCE loss assigns higher weights to the harder negative examples to improve the fine-grained feature representation of the model. * Bidirectional-attention: Remove the causal attention of LLMs during contrastive training of decoder-only LLM to produce rich contextualized representations. * Training efficiency: Using Gradient Cache to scale contrastive learning batches beyond GPU memory constraints allows the model to learn from more challenging negative examples. * Others： Dataset-Homogenous Batching、cross-batch negative sampling ### Model Details * Base Decoder-only LLM: [gte-Qwen2-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct) * Pooling Methods: Last token * Embedding Dimension: 3584 ### Usage ##### Requirements ``` transformers>=4.40.2 flash_attn>=2.5.8 sentence-transformers>=2.7.0 ``` ##### How to use Here is an example of how to encode queries and passages using Huggingface-transformer and Sentence-transformer. ##### Usage (HuggingFace Transformers) ```python import torch import torch.nn.functional as F from torch import Tensor from transformers import AutoTokenizer, AutoModel def last_token_pool(last_hidden_states: Tensor, attention_mask: Tensor) -> Tensor: left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0]) if left_padding: return last_hidden_states[:, -1] else: sequence_lengths = attention_mask.sum(dim=1) - 1 batch_size = last_hidden_states.shape[0] return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths] def get_detailed_instruct(task_description: str, query: str) -> str: return f'Instruct: {task_description}\nQuery: {query}' task = 'Given a web search query, retrieve relevant passages that answer the query' queries = [ get_detailed_instruct(task, "国家法定节假日共多少天"), get_detailed_instruct(task, "如何查看好友申请") ] documents = [ "一年国家法定节假日为11天。根据公布的国家法定节假日调整方案，调整的主要内容包括：元旦放假1天不变；春节放假3天，放假时间为农历正月初一、初二、初三；“五一”国际劳动节1天不变；“十一”国庆节放假3天；清明节、端午节、中秋节增设为国家法定节假日，各放假1天(农历节日如遇闰月，以第一个月为休假日)。3、允许周末上移下错，与法定节假日形成连休。", "这个直接去我的QQ中心不就好了么那里可以查到 我的好友单向好友好友恢复、 以及好友申请 啊可以是你加别人的 或 别人加你的都可以查得到QQ空间里 这个没注意 要有的话也会在你进空间的时候会提示你的QQ 空间里 上面消息 就可以看见了！望采纳！谢谢这个直接去我的QQ中心不就好了么那里可以查到 我的好友单向好友好友恢复、 以及好友申请 啊可以是你加别人的 或 别人加你的都可以查得到", ] input_texts = queries + documents tokenizer = AutoTokenizer.from_pretrained('Lenovo-Zhihui/Zhihui_LLM_Embedding', trust_remote_code=True) model = AutoModel.from_pretrained('Lenovo-Zhihui/Zhihui_LLM_Embedding', trust_remote_code=True) max_length = 512 # Tokenize the input texts batch_dict = tokenizer(input_texts, max_length=max_length, padding=True, truncation=True, return_tensors='pt') outputs = model(**batch_dict) embeddings = last_token_pool(outputs.last_hidden_state, batch_dict['attention_mask']) # normalize embeddings embeddings = F.normalize(embeddings, p=2, dim=1) scores = (embeddings[:2] @ embeddings[2:].T) print(scores.tolist()) ``` ##### Usage (Sentence-Transformers) ```python from sentence_transformers import SentenceTransformer model = SentenceTransformer("Lenovo-Zhihui/Zhihui_LLM_Embedding", trust_remote_code=True) model.max_seq_length = 512 # 数据来源DuRetrieval https://huggingface.co/datasets/C-MTEB/DuRetrieval queries = [ "国家法定节假日共多少天", "如何查看好友申请", ] documents = [ "一年国家法定节假日为11天。根据公布的国家法定节假日调整方案，调整的主要内容包括：元旦放假1天不变；春节放假3天，放假时间为农历正月初一、初二、初三；“五一”国际劳动节1天不变；“十一”国庆节放假3天；清明节、端午节、中秋节增设为国家法定节假日，各放假1天(农历节日如遇闰月，以第一个月为休假日)。3、允许周末上移下错，与法定节假日形成连休。", "这个直接去我的QQ中心不就好了么那里可以查到 我的好友单向好友好友恢复、 以及好友申请 啊可以是你加别人的 或 别人加你的都可以查得到QQ空间里 这个没注意 要有的话也会在你进空间的时候会提示你的QQ 空间里 上面消息 就可以看见了！望采纳！谢谢这个直接去我的QQ中心不就好了么那里可以查到 我的好友单向好友好友恢复、 以及好友申请 啊可以是你加别人的 或 别人加你的都可以查得到", ] query_embeddings = model.encode(queries, prompt_name="query", normalize_embeddings=True) document_embeddings = model.encode(documents, normalize_embeddings=True) scores = (query_embeddings @ document_embeddings.T) print(scores.tolist()) ``` ### Reproduce our results（C-MTEB）: Check out scripts/eval_mteb.py to reproduce evaluation results on C-MTEB benchmark. | Model | T2Retrieval | MMarcoRetrieval | DuRetrieval | CovidRetrieval | CmedqaRetrieval | EcomRetrieval | MedicalRetrieval | VideoRetrieval | Avg | |:-------------------------------|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:|:--------:| |**Zhihui_LLM_Embedding** | 88.30 | 84.77 | 91.34 | 84.39 | 48.69 | 71.96 | 65.19 | 79.31 | **76.74** | |zpoint_large_embedding_zh | 83.81 | 82.38 | 89.23 | 89.14 | 47.16 | 70.74 | 68.14 | 80.26 | 76.36 | |gte-Qwen2-7B-instruct | 87.73 | 85.16 | 87.44 | 83.65 | 48.69 | 71.15 | 65.59 | 78.84 | 76.03 | |360Zhinao-search | 87.12 | 83.32 | 87.57 | 85.02 | 46.73 | 68.9 | 63.69 | 78.09 | 75.06 | |AGE_Hybrid | 86.88 | 80.65 | 89.28 | 83.66 | 47.26 | 69.28 | 65.94 | 76.79 | 74.97 |