Hugging Face's logo

infly
/
inf-retriever-v1-1.5b

Sentence Similarity
sentence-transformers
Safetensors
Transformers
English
Chinese
qwen2
feature-extraction
custom_code
text-generation-inference
text-embeddings-inference
Model card Files
xet
Community
3
SamuelYang commited on
Commit
f3f60cc
·
verified ·
1 Parent(s): c9c05c2

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +179 -0
README.md ADDED
@@ -0,0 +1,179 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model:
3
+ - Alibaba-NLP/gte-Qwen2-1.5B-instruct
4
+ language:
5
+ - en
6
+ - zh
7
+ license: apache-2.0
8
+ tags:
9
+ - sentence-transformers
10
+ - transformers
11
+ - sentence-similarity
12
+ ---
13
+ # INF-Retriever-v1-1.5B
14
+ ## Model Overview
15
+ - **INF-Retriever-v1-1.5B** is a lightweight version of the [**INF-Retriever-v1**](https://huggingface.co/infly/inf-retriever-v1), an LLM-based dense retrieval model developed by [INF TECH](https://www.infly.cn/en).
16
+ It is built upon the [gte-Qwen2-1.5B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-1.5B-instruct) model and specifically fine-tuned to excel in retrieval tasks, particularly for Chinese and English data.
17
+
18
+ - As of February 19, 2025, **INF-Retriever-v1-1.5B** ranks both **No.1** on the Automated Heterogeneous Information Retrieval Benchmark of version 24.04 & 24.05([AIR-Bench](https://huggingface.co/spaces/AIR-Bench/leaderboard)) for the bilingual Chinese and English sub-leaderboard, among models with fewer than 7B parameters. This demonstrates its cutting-edge performance in heterogeneous information retrieval tasks.
19
+
20
+ ## Key Features
21
+
22
+ - **Optimized for Chinese and English retrieval**: The model has been specifically fine-tuned with retrieval-focused datasets in both languages, significantly improving its accuracy and efficiency for a variety of retrieval scenarios.
23
+
24
+ - **Top-tier performance**: **INF-Retriever-v1-1.5B** has achieved outstanding results on the AIR-Bench leaderboard, making it a top choice for heterogeneous information retrieval tasks across various domains.
25
+
26
+ ## Model Details
27
+ - Model Size: 1.5B
28
+ - Embedding Dimension: 1536
29
+ - Max Input Tokens: 32768
30
+ - Language Support: Chinese & English (also effective in other languages)
31
+
32
+ ## Usage
33
+
34
+ ### Sentence Transformers
35
+ ```python
36
+ from sentence_transformers import SentenceTransformer
37
+
38
+ model = SentenceTransformer("infly/inf-retriever-v1-1.5b", trust_remote_code=True)
39
+ # In case you want to reduce the maximum length:
40
+ model.max_seq_length = 8192
41
+
42
+ queries = [
43
+ "how much protein should a female eat",
44
+ "summit define",
45
+ ]
46
+ documents = [
47
+ "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
48
+ "Definition of summit for English Language Learners. : 1 the highest point of a mountain : the top of a mountain. : 2 the highest level. : 3 a meeting or series of meetings between the leaders of two or more governments.",
49
+ ]
50
+
51
+ query_embeddings = model.encode(queries, prompt_name="query")
52
+ document_embeddings = model.encode(documents)
53
+
54
+ scores = (query_embeddings @ document_embeddings.T) * 100
55
+ print(scores.tolist())
56
+ # [[89.36092376708984, 69.16694641113281], [57.51953125, 79.65923309326172]]
57
+ ```
58
+
59
+ ### Transformers
60
+ ```python
61
+ import torch
62
+ import torch.nn.functional as F
63
+
64
+ from torch import Tensor
65
+ from transformers import AutoTokenizer, AutoModel
66
+
67
+
68
+ def last_token_pool(last_hidden_states: Tensor,
69
+ attention_mask: Tensor) -> Tensor:
70
+ left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])
71
+ if left_padding:
72
+ return last_hidden_states[:, -1]
73
+ else:
74
+ sequence_lengths = attention_mask.sum(dim=1) - 1
75
+ batch_size = last_hidden_states.shape[0]
76
+ return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]
77
+
78
+
79
+ def get_detailed_instruct(task_description: str, query: str) -> str:
80
+ return f'Instruct: {task_description}\nQuery: {query}'
81
+
82
+
83
+ # Each query must come with a one-sentence instruction that describes the task
84
+ task = 'Given a web search query, retrieve relevant passages that answer the query'
85
+ queries = [
86
+ get_detailed_instruct(task, 'how much protein should a female eat'),
87
+ get_detailed_instruct(task, 'summit define')
88
+ ]
89
+ # No need to add instruction for retrieval documents
90
+ documents = [
91
+ "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
92
+ "Definition of summit for English Language Learners. : 1 the highest point of a mountain : the top of a mountain. : 2 the highest level. : 3 a meeting or series of meetings between the leaders of two or more governments."
93
+ ]
94
+ input_texts = queries + documents
95
+
96
+ tokenizer = AutoTokenizer.from_pretrained('infly/inf-retriever-v1-1.5b', trust_remote_code=True)
97
+ model = AutoModel.from_pretrained('infly/inf-retriever-v1-1.5b', trust_remote_code=True)
98
+
99
+ max_length = 8192
100
+
101
+ # Tokenize the input texts
102
+ batch_dict = tokenizer(input_texts, max_length=max_length, padding=True, truncation=True, return_tensors='pt')
103
+ outputs = model(**batch_dict)
104
+ embeddings = last_token_pool(outputs.last_hidden_state, batch_dict['attention_mask'])
105
+
106
+ # normalize embeddings
107
+ embeddings = F.normalize(embeddings, p=2, dim=1)
108
+ scores = (embeddings[:2] @ embeddings[2:].T) * 100
109
+ print(scores.tolist())
110
+ # [[89.36091613769531, 69.16694641113281], [57.519447326660156, 79.65917205810547]]
111
+ ```
112
+
113
+ ## Evaluation
114
+
115
+ ### AIR-Bench
116
+
117
+ **INF-Retriever-v1-1.5B** has demonstrated superior retrieval capabilities across multiple domains and languages. The results from the Automated Heterogeneous Information Retrieval Benchmark ([AIR-Bench](https://huggingface.co/spaces/AIR-Bench/leaderboard)) as of February 19, 2025, are as follows:
118
+
119
+ #### AIR-Bench_24.04 (Bilingual, EN & ZH)
120
+
121
+ | Model Name | Under 7B | Average⬆️ | wiki_en | wiki_zh | web_en | web_zh | healthcare_en | healthcare_zh | law_en | arxiv_en | news_en | news_zh | finance_en | finance_zh | msmarco_en |
122
+ |---------------------------------------------------------------------------------------|:--------:|-----------|-----------|-----------|-----------|----------|---------------|---------------|-----------|-----------|-----------|-----------|------------|------------|------------|
123
+ | [GTE-Qwen1.5-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen1.5-7B-instruct) | ❌ | 41.61 | 57.05 | 52.89 | 43.17 | 44.9 | 54.44 | 37.42 | 11.85 | 32.31 | 50.07 | 24.19 | 55.16 | 26.09 | 51.35 |
124
+ | [Multilingual-E5-large](https://huggingface.co/intfloat/multilingual-e5-large) | ✅ | 42.58 | 53.76 | 60.57 | 37.55 | 48.27 | 50.63 | 33.74 | 19.66 | 36.93 | 43.5 | 39.72 | 47.77 | 26.98 | 54.44 |
125
+ | [E5-mistral-7b-instruct](https://huggingface.co/intfloat/e5-mistral-7b-instruct) | ❌ | 45.26 | 61.67 | 55.97 | 44.41 | 45.96 | 56.32 | 35.79 | 19.32 | 44.78 | 48.18 | 35.99 | 54.79 | 26.11 | 59.03 |
126
+ | [BGE-M3](https://huggingface.co/BAAI/bge-m3) | ✅ | 46.65 | 60.49 | 62.36 | 47.35 | 50.38 | 49.1 | **42.38** | 26.68 | 40.76 | 48.04 | 40.75 | 51.52 | 32.18 | 54.4 |
127
+ | [BGE-Multilingual-Gemma2](https://huggingface.co/BAAI/bge-multilingual-gemma2) | ❌ | 46.83 | 63.71 | 67.3 | 50.38 | 53.24 | 47.24 | 42.13 | 22.58 | 23.28 | 50.91 | 44.02 | 49.3 | 31.6 | **63.14** |
128
+ | [GTE-Qwen2-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct) | ❌ | 48.38 | 63.46 | 66.44 | 51.2 | 51.98 | 54.2 | 38.82 | 22.31 | 40.27 | **54.07** | 43.03 | 58.2 | 26.63 | 58.39 |
129
+ | **INF-Retriever-v1-1.5B** | ✅ | 49.77 | 62.87 | 65.98 | 50.16 | 53.8 | 54.48 | 40.22 | 32 | 45.3 | 51.47 | 46.02 | 56.81 | 31.15 | 56.73 |
130
+ | [INF-Retriever-v1](https://huggingface.co/infly/inf-retriever-v1) | ❌ | **52.56** | **65.25** | **68.44** | **52.13** | **56.6** | **56.96** | 42.03 | **34.51** | **50.62** | 53.32 | **50.02** | **58.34** | **35.42** | 59.64 |
131
+
132
+ #### AIR-Bench_24.05 (Multilingual, 13 languages)
133
+
134
+ ##### Bilingual (EN & ZH)
135
+
136
+ | Model Name | Under 7B | Average⬆️ | wiki_en | wiki_zh | web_en | web_zh | healthcare_en | healthcare_zh | law_en | arxiv_en | news_en | news_zh | finance_en | finance_zh |
137
+ |---------------------------------------------------------------------------------------|:--------:|-----------|-----------|-----------|-----------|-----------|---------------|---------------|-----------|-----------|-----------|-----------|------------|------------|
138
+ | [GTE-multilingual-base](https://huggingface.co/Alibaba-NLP/gte-multilingual-base) | ✅ | 45.14 | 69.12 | 61.86 | 52.05 | 46.75 | 47.48 | 37.94 | 11.44 | 41.28 | 47.54 | 36.2 | 53.24 | 36.84 |
139
+ | [GTE-Qwen1.5-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen1.5-7B-instruct) | ❌ | 45.99 | 66.45 | 58.33 | 52.68 | 47.48 | 52.11 | 39.13 | 20.19 | 42.15 | 47.44 | 36.43 | 55.21 | 34.28 |
140
+ | [E5-mistral-7b-instruct](https://huggingface.co/intfloat/e5-mistral-7b-instruct) | ❌ | 46.43 | 71.38 | 57.19 | 52.08 | 45.68 | 56.24 | 36.05 | 19.61 | 46.06 | 47.89 | 35.98 | 55.9 | 33.1 |
141
+ | [BGE-Multilingual-Gemma2](https://huggingface.co/BAAI/bge-multilingual-gemma2) | ❌ | 47.53 | 72.8 | 68.64 | 56.48 | 53.04 | 47.48 | **42.35** | 22.6 | 24 | 50.29 | 43.42 | 50.08 | 39.23 |
142
+ | [BGE-M3](https://huggingface.co/BAAI/bge-m3) | ✅ | 48.23 | 69.7 | 63.52 | 53.88 | 50.2 | 49.05 | 42.31 | 26.95 | 41.64 | 47.34 | 41 | 52.92 | 40.23 |
143
+ | [GTE-Qwen2-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct) | ❌ | 49.89 | **73.59** | 67.5 | **58.99** | 51.66 | 54.46 | 38.66 | 22.75 | 41.32 | **52.74** | 43.17 | 59.23 | 34.61 |
144
+ | **INF-Retriever-v1-1.5B** | ✅ | 51.28 | 71.58 | 67.04 | 55.93 | 53.23 | 54.72 | 40.35 | 32.37 | 46.34 | 50.66 | 45.7 | 58.08 | 39.37 |
145
+ | [INF-Retriever-v1](https://huggingface.co/infly/inf-retriever-v1) | ❌ | **54.01** | 73.52 | **69.45** | 57.6 | **56.46** | **57.03** | 41.82 | **34.76** | **51.38** | 52.7 | **49.78** | **59.44** | **44.13** |
146
+
147
+ ##### Multilingual (13 languages)
148
+ Although INF-Retriever-v1-1.5B has been fine-tuned exclusively on English and Chinese, it continues to perform exceptionally well across other languages.
149
+
150
+ | Model Name | Under 7B | Average⬆️ | wiki_en | wiki_zh | wiki_ar | wiki_bn | wiki_de | wiki_es | wiki_fa | wiki_fr | wiki_hi | wiki_id | wiki_ja | wiki_ko | wiki_ru | web_en | web_zh | web_ar | web_bn | web_de | web_es | web_fa | web_fr | web_hi | web_id | web_ja | web_ko | web_ru | healthcare_en | healthcare_zh | healthcare_de | healthcare_es | healthcare_fr | law_en | law_de | law_fr | arxiv_en | science_ru | news_en | news_zh | news_ar | news_bn | news_de | news_es | news_fa | news_fr | news_hi | news_id | news_ja | news_ko | news_ru | finance_en | finance_zh | finance_ar | finance_fr |
151
+ |--------------------------------------------------------------------------------------------------|:--------:|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|----------|-----------|-----------|----------|--------|-----------|-----------|-----------|---------------|---------------|---------------|---------------|---------------|-----------|-----------|-----------|-----------|------------|-----------|-----------|-----------|-----------|-----------|----------|-----------|----------|-----------|-----------|-----------|-----------|-----------|------------|------------|------------|------------|
152
+ | [E5-mistral-7b-instruct](https://huggingface.co/intfloat/e5-mistral-7b-instruct) | ❌ | 48.08 | 71.38 | 57.19 | 52.98 | 56.84 | 65.4 | 69.49 | 51.77 | 69.29 | 63.93 | 66.23 | 57.72 | 60.3 | 58.7 | 52.08 | 45.68 | 49.56 | 46.83 | 50.88 | 54.46 | 45.86 | 54.52 | 49.43 | 55.17 | 51.8 | 54.22 | 53.85 | 56.24 | 36.05 | 53.12 | 47.67 | 37.28 | 19.61 | 14.77 | 14.38 | 46.06 | 53.07 | 47.89 | 35.98 | 38.95 | 25.5 | 46.48 | 45.34 | 29.72 | 49.61 | 29.82 | 45.93 | 43.47 | 46.46 | 46.59 | 55.9 | 33.1 | 44.59 | 38.98 |
153
+ | [jina-embeddings-v3](https://huggingface.co/jinaai/jina-embeddings-v3) | ✅ | 48.46 | 64.96 | 62.7 | 57.89 | 62.81 | 62.08 | 63.65 | 57.75 | 64.67 | 68.74 | 62.75 | 58.26 | 58.28 | 59.41 | 47.38 | 47.66 | 53.4 | 55.55 | 48.06 | 49.42 | 52.84 | 48.8 | 58.79 | 52.76 | 50.1 | 51.87 | 50.51 | 49.42 | 38.92 | 49.86 | 52.75 | 32.68 | 16.78 | 11.71 | 9.76 | 39.65 | 50.24 | 45.61 | 40.56 | 44.04 | 53.73 | 46.39 | 42.94 | 37.9 | 46.56 | 40.02 | 44.86 | 41.96 | 45.18 | 46.65 | 51.7 | 33.96 | 46.32 | 37.14 |
154
+ | **INF-Retriever-v1-1.5B** | ✅ | 50 | 71.58 | 67.04 | 59.44 | 56.53 | 64.11 | 67.57 | 57.75 | 68.12 | 63.86 | 64.64 | 62.02 | 63.43 | 60.6 | 55.93 | 53.23 | 52.7 | 43.52 | 50.65 | 52.97 | 47.64 | 53.76 | 43.05 | 54.55 | 56.95 | 56.49 | 55.05 | 54.72 | 40.35 | 48.68 | 54.29 | 39.28 | 32.37 | 18.12 | 17.79 | 46.34 | 54.7 | 50.66 | 45.7 | 43.84 | 24.33 | 47.72 | 43.8 | 32.64 | 51.49 | 27.05 | 44.49 | 47.62 | 49.3 | 47.59 | 58.08 | 39.37 | 45.99 | 40.57 |
155
+ | [GTE-Qwen2-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct) | ❌ | 50.05 | **73.59** | 67.5 | 59.44 | 58.17 | 63.96 | 67.62 | 57.05 | 70.32 | 60.54 | 61.81 | 62.88 | 59.17 | 62.95 | **58.99** | 51.66 | 55.56 | 51.45 | 48.62 | 54.11 | 49.54 | 55.16 | 53.06 | 55.51 | 57.27 | 57.54 | 55.88 | 54.46 | 38.66 | 53.92 | 53.78 | 30.29 | 22.75 | 13.18 | 13.15 | 41.32 | 45.21 | **52.74** | 43.17 | 37.63 | **61.31** | 44.89 | 45.21 | 30.1 | 49.76 | 30.28 | 46.44 | 44.13 | 47.19 | 46.55 | 59.23 | 34.61 | 43.56 | 39.57 |
156
+ | [Multilingual-E5-large-instruct](https://huggingface.co/intfloat/multilingual-e5-large-instruct) | ✅ | 51.11 | 68.62 | 62.82 | 63.21 | 64.45 | 65.81 | 68.1 | 64.2 | 69.72 | 71.81 | 66.36 | 64.12 | 64.79 | 62.57 | 41.58 | 47.06 | 56.4 | 56.17 | 50.87 | 52.24 | 58.68 | 50.2 | 56.32 | 54.49 | 54.89 | 55.81 | 54.97 | 54.02 | 39.76 | 52.06 | 51.74 | 36.64 | 16.9 | 15.59 | 15.12 | 39.52 | 56.86 | 44.28 | 35.46 | 48.2 | 49.31 | 47.84 | 45.99 | **45.59** | 50.58 | 39.66 | 48.59 | 47.6 | 50.52 | 48.81 | 52.79 | 37.72 | 48.95 | 42.74 |
157
+ | [BGE-M3](https://huggingface.co/BAAI/bge-m3) | ✅ | 51.31 | 69.7 | 63.52 | 59.65 | 64.33 | 64.68 | 65.4 | 61.14 | 66.04 | 69.02 | 66.3 | 60.86 | 62.36 | 60.18 | 53.88 | 50.2 | 52.53 | 55.53 | 51.89 | 51.78 | 55.81 | 51.46 | 57.06 | 53.14 | 54.75 | 55.28 | 54.53 | 49.05 | 42.31 | 49 | 53.05 | 39.29 | 26.95 | 20.11 | 20.2 | 41.64 | 55.18 | 47.34 | 41 | 44.93 | 59.03 | 47.87 | 44.7 | 43.81 | 49.52 | 42.12 | 47.45 | 47.09 | 48.14 | 48.31 | 52.92 | 40.23 | 45.76 | 41.44 |
158
+ | [BGE-Multilingual-Gemma2](https://huggingface.co/BAAI/bge-multilingual-gemma2) | ❌ | 54.46 | 72.8 | 68.64 | **63.42** | **69.48** | **67.91** | **71.79** | **67.57** | **71.28** | **75.39** | **68.91** | **68.29** | **66.78** | **64.15** | 56.48 | 53.04 | **59.97** | **59.68** | **57.72** | **58.2** | **62.43** | **59.54** | **64.5** | **60** | **60.26** | 59.64 | **60.12** | 47.48 | **42.35** | 55.4 | **63.13** | **45.13** | 22.6 | 15.75 | 14.29 | 24 | 44.13 | 50.29 | 43.42 | 48.41 | 58.77 | **52.05** | **49.9** | 43.4 | **56.8** | **44.89** | 50.65 | **51.51** | 51.64 | 51.48 | 50.08 | 39.23 | 50.25 | **51.1** |
159
+ | [INF-Retriever-v1](https://huggingface.co/infly/inf-retriever-v1) | ❌ | **54.47** | 73.52 | **69.45** | 63.13 | 61.58 | 66.8 | 69.29 | 63.03 | 69.74 | 69.02 | 68.63 | 63.45 | 64.44 | 62.74 | 57.6 | **56.46** | 58.48 | 53.7 | 55.2 | 57.08 | 53.27 | 57.35 | 55.64 | 58.85 | 59.52 | **60.01** | 58.79 | **57.03** | 41.82 | **55.46** | 57.6 | 43.25 | **34.76** | **21.75** | **21.87** | **51.38** | **59.72** | 52.7 | **49.78** | **49.11** | 43.62 | 51.47 | 49.52 | 40.43 | 54.54 | 38.57 | **51.06** | 51.12 | **53.15** | **51.88** | **59.44** | **44.13** | **50.71** | 44.2 |
160
+
161
+ ## Contributors
162
+ ### Supervisors
163
+ Wei Chu • Yinghui Xu • Yuan Qi
164
+ ### INF memory team
165
+ Junhan Yang ([email protected]) • Jiahe Wan • Yichen Yao ([email protected])
166
+
167
+ ## Citation
168
+ If you find our model useful, please consider citing:
169
+
170
+ ```
171
+ @misc {infly-ai_2025,
172
+ author = { {infly-ai} },
173
+ title = { inf-retriever-v1 (Revision 5f469d7) },
174
+ year = 2025,
175
+ url = { https://huggingface.co/infly/inf-retriever-v1 },
176
+ doi = { 10.57967/hf/4262 },
177
+ publisher = { Hugging Face }
178
+ }
179
+ ```