metadata
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
- 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)
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)
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 |