nreimers commited on
Commit
7c6bc2e
1 Parent(s): f48dad6
README.md ADDED
@@ -0,0 +1,101 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Sentence Embedding Model for MS MARCO Passage Retrieval
2
+
3
+
4
+ This a `distilroberta-base` model from the [sentence-transformers](https://github.com/UKPLab/sentence-transformers)-repository. It was trained on the [MS MARCO Passage Retrieval dataset](https://github.com/microsoft/MSMARCO-Passage-Ranking): Given a search query, it finds the relevant passages.
5
+
6
+ You can use this model for semantic search. Details can be found on: [SBERT.net - Semantic Search](https://www.sbert.net/examples/applications/semantic-search/README.html).
7
+
8
+ This model was optimized to be used with **cosine-similarity** as similarity function between queries and documents.
9
+
10
+
11
+ ## Training
12
+
13
+ Details about the training of the models can be found here: [SBERT.net - MS MARCO](https://www.sbert.net/examples/training/ms_marco/README.html)
14
+
15
+ ## Performance
16
+
17
+ For performance details, see: [SBERT.net - Pre-Trained Models - MS MARCO](https://www.sbert.net/docs/pretrained-models/msmarco-v3.html)
18
+
19
+ ## Usage (HuggingFace Models Repository)
20
+
21
+ You can use the model directly from the model repository to compute sentence embeddings:
22
+ ```python
23
+ from transformers import AutoTokenizer, AutoModel
24
+ import torch
25
+
26
+
27
+ #Mean Pooling - Take attention mask into account for correct averaging
28
+ def mean_pooling(model_output, attention_mask):
29
+ token_embeddings = model_output[0] #First element of model_output contains all token embeddings
30
+ input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
31
+ sum_embeddings = torch.sum(token_embeddings * input_mask_expanded, 1)
32
+ sum_mask = torch.clamp(input_mask_expanded.sum(1), min=1e-9)
33
+ return sum_embeddings / sum_mask
34
+
35
+
36
+ # Queries we want embeddings for
37
+ queries = ['What is the capital of France?', 'How many people live in New York City?']
38
+
39
+ # Passages that provide answers
40
+ passages = ['Paris is the capital of France', 'New York City is the most populous city in the United States, with an estimated 8,336,817 people living in the city, according to U.S. Census estimates dating July 1, 2019']
41
+
42
+ #Load AutoModel from huggingface model repository
43
+ tokenizer = AutoTokenizer.from_pretrained("model_name")
44
+ model = AutoModel.from_pretrained("model_name")
45
+
46
+ def compute_embeddings(sentences):
47
+ #Tokenize sentences
48
+ encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
49
+
50
+ #Compute query embeddings
51
+ with torch.no_grad():
52
+ model_output = model(**encoded_input)
53
+
54
+ #Perform pooling. In this case, mean pooling
55
+ return mean_pooling(model_output, encoded_input['attention_mask'])
56
+
57
+ query_embeddings = compute_embeddings(queries)
58
+ passage_embeddings = compute_embeddings(passages)
59
+ ```
60
+
61
+ ## Usage (Sentence-Transformers)
62
+ Using this model becomes more convenient when you have [sentence-transformers](https://github.com/UKPLab/sentence-transformers) installed:
63
+ ```
64
+ pip install -U sentence-transformers
65
+ ```
66
+
67
+ Then you can use the model like this:
68
+ ```python
69
+ from sentence_transformers import SentenceTransformer
70
+ model = SentenceTransformer('model_name')
71
+
72
+ # Queries we want embeddings for
73
+ queries = ['What is the capital of France?', 'How many people live in New York City?']
74
+
75
+ # Passages that provide answers
76
+ passages = ['Paris is the capital of France', 'New York City is the most populous city in the United States, with an estimated 8,336,817 people living in the city, according to U.S. Census estimates dating July 1, 2019']
77
+
78
+ query_embeddings = model.encode(queries)
79
+ passage_embeddings = model.encode(passages)
80
+ ```
81
+
82
+ ## Changes in v3
83
+ The models from v2 have been used for find for all training queries similar passages. An [MS MARCO Cross-Encoder](ce-msmarco.md) based on the electra-base-model has been then used to classify if these retrieved passages answer the question.
84
+
85
+ If they received a low score by the cross-encoder, we saved them as hard negatives: They got a high score from the bi-encoder, but a low-score from the (better) cross-encoder.
86
+
87
+ We then trained the v2 models with these new hard negatives.
88
+
89
+ ## Citing & Authors
90
+ If you find this model helpful, feel free to cite our publication [Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks](https://arxiv.org/abs/1908.10084):
91
+ ```
92
+ @inproceedings{reimers-2019-sentence-bert,
93
+ title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
94
+ author = "Reimers, Nils and Gurevych, Iryna",
95
+ booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
96
+ month = "11",
97
+ year = "2019",
98
+ publisher = "Association for Computational Linguistics",
99
+ url = "http://arxiv.org/abs/1908.10084",
100
+ }
101
+ ```
config.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "../output/microsoft_MiniLM-L12-H384-uncased-mined_hard_neg-mean-pooling-no_identifier-epoch10-batchsize70-2021-04-05_07-37-55/0_Transformer",
3
+ "architectures": [
4
+ "BertModel"
5
+ ],
6
+ "attention_probs_dropout_prob": 0.1,
7
+ "gradient_checkpointing": false,
8
+ "hidden_act": "gelu",
9
+ "hidden_dropout_prob": 0.1,
10
+ "hidden_size": 384,
11
+ "initializer_range": 0.02,
12
+ "intermediate_size": 1536,
13
+ "layer_norm_eps": 1e-12,
14
+ "max_position_embeddings": 512,
15
+ "model_type": "bert",
16
+ "num_attention_heads": 12,
17
+ "num_hidden_layers": 12,
18
+ "pad_token_id": 0,
19
+ "position_embedding_type": "absolute",
20
+ "transformers_version": "4.4.2",
21
+ "type_vocab_size": 2,
22
+ "use_cache": true,
23
+ "vocab_size": 30522
24
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:57555aab9e3b37aa1e185d6e4534fe279867ec793aebf9e5176294dcb2519114
3
+ size 133526519
sentence_bert_config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "max_seq_length": 512,
3
+ "do_lower_case": false
4
+ }
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]"}
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"do_lower_case": true, "unk_token": "[UNK]", "sep_token": "[SEP]", "pad_token": "[PAD]", "cls_token": "[CLS]", "mask_token": "[MASK]", "tokenize_chinese_chars": true, "strip_accents": null, "model_max_length": 512, "name_or_path": "microsoft/MiniLM-L12-H384-uncased", "do_basic_tokenize": true, "never_split": null}
vocab.txt ADDED
The diff for this file is too large to render. See raw diff