espejelomar commited on
Commit
7ba6e49
1 Parent(s): caba8d1

Add new SentenceTransformer model.

Browse files
.gitattributes CHANGED
@@ -25,3 +25,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
25
  *.zip filter=lfs diff=lfs merge=lfs -text
26
  *.zstandard filter=lfs diff=lfs merge=lfs -text
27
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
25
  *.zip filter=lfs diff=lfs merge=lfs -text
26
  *.zstandard filter=lfs diff=lfs merge=lfs -text
27
  *tfevents* filter=lfs diff=lfs merge=lfs -text
28
+ pytorch_model.bin filter=lfs diff=lfs merge=lfs -text
1_Pooling/config.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
1
+ {
2
+ "word_embedding_dimension": 768,
3
+ "pooling_mode_cls_token": false,
4
+ "pooling_mode_mean_tokens": true,
5
+ "pooling_mode_max_tokens": false,
6
+ "pooling_mode_mean_sqrt_len_tokens": false
7
+ }
README.md ADDED
@@ -0,0 +1,177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ pipeline_tag: sentence-similarity
3
+ tags:
4
+ - sentence-transformers
5
+ - feature-extraction
6
+ - sentence-similarity
7
+ language:
8
+ - es
9
+ datasets:
10
+ - hackathon-pln-es/nli-es
11
+ widget:
12
+ - text: "A ver si nos tenemos que poner todos en huelga hasta cobrar lo que queramos."
13
+ - text: "La huelga es el método de lucha más eficaz para conseguir mejoras en el salario."
14
+ - text: "Tendremos que optar por hacer una huelga para cobrar lo que queremos."
15
+ - text: "Queda descartada la huelga aunque no cobremos lo que queramos."
16
+ ---
17
+
18
+ # bertin-roberta-base-finetuning-esnli
19
+
20
+ This is a [sentence-transformers](https://www.SBERT.net) model trained on a
21
+ collection of NLI tasks for Spanish. It maps sentences & paragraphs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search.
22
+
23
+ Based around the siamese networks approach from [this paper](https://arxiv.org/pdf/1908.10084.pdf).
24
+ <!--- Describe your model here -->
25
+
26
+ You can see a demo for this model [here](https://huggingface.co/spaces/hackathon-pln-es/Sentence-Embedding-Bertin).
27
+
28
+ You can find our other model, **paraphrase-spanish-distilroberta** [here](https://huggingface.co/hackathon-pln-es/paraphrase-spanish-distilroberta) and its demo [here](https://huggingface.co/spaces/hackathon-pln-es/Paraphrase-Bertin).
29
+
30
+ ## Usage (Sentence-Transformers)
31
+
32
+ Using this model becomes easy when you have [sentence-transformers](https://www.SBERT.net) installed:
33
+
34
+ ```
35
+ pip install -U sentence-transformers
36
+ ```
37
+
38
+ Then you can use the model like this:
39
+
40
+ ```python
41
+ from sentence_transformers import SentenceTransformer
42
+ sentences = ["Este es un ejemplo", "Cada oración es transformada"]
43
+
44
+ model = SentenceTransformer('hackathon-pln-es/bertin-roberta-base-finetuning-esnli')
45
+ embeddings = model.encode(sentences)
46
+ print(embeddings)
47
+ ```
48
+
49
+ ## Usage (HuggingFace Transformers)
50
+ Without [sentence-transformers](https://www.SBERT.net), you can use the model like this: First, you pass your input through the transformer model, then you have to apply the right pooling-operation on-top of the contextualized word embeddings.
51
+
52
+ ```python
53
+ from transformers import AutoTokenizer, AutoModel
54
+ import torch
55
+
56
+
57
+ #Mean Pooling - Take attention mask into account for correct averaging
58
+ def mean_pooling(model_output, attention_mask):
59
+ token_embeddings = model_output[0] #First element of model_output contains all token embeddings
60
+ input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
61
+ return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)
62
+
63
+
64
+ # Sentences we want sentence embeddings for
65
+ sentences = ['This is an example sentence', 'Each sentence is converted']
66
+
67
+ # Load model from HuggingFace Hub
68
+ tokenizer = AutoTokenizer.from_pretrained('hackathon-pln-es/bertin-roberta-base-finetuning-esnli')
69
+ model = AutoModel.from_pretrained('hackathon-pln-es/bertin-roberta-base-finetuning-esnli')
70
+
71
+ # Tokenize sentences
72
+ encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')
73
+
74
+ # Compute token embeddings
75
+ with torch.no_grad():
76
+ model_output = model(**encoded_input)
77
+
78
+ # Perform pooling. In this case, mean pooling.
79
+ sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
80
+
81
+ print("Sentence embeddings:")
82
+ print(sentence_embeddings)
83
+ ```
84
+
85
+
86
+ ## Evaluation Results
87
+
88
+ <!--- Describe how your model was evaluated -->
89
+ Our model was evaluated on the task of Semantic Textual Similarity using the [SemEval-2015 Task](https://alt.qcri.org/semeval2015/task2/) for [Spanish](http://alt.qcri.org/semeval2015/task2/data/uploads/sts2015-es-test.zip). We measure
90
+
91
+ | | [BETO STS](https://huggingface.co/espejelomar/sentece-embeddings-BETO) | BERTIN STS (this model) | Relative improvement |
92
+ |-------------------:|---------:|-----------:|---------------------:|
93
+ | cosine_pearson | 0.609803 | 0.683188 | +12.03 |
94
+ | cosine_spearman | 0.528776 | 0.615916 | +16.48 |
95
+ | euclidean_pearson | 0.590613 | 0.672601 | +13.88 |
96
+ | euclidean_spearman | 0.526529 | 0.611539 | +16.15 |
97
+ | manhattan_pearson | 0.589108 | 0.672040 | +14.08 |
98
+ | manhattan_spearman | 0.525910 | 0.610517 | +16.09 |
99
+ | dot_pearson | 0.544078 | 0.600517 | +10.37 |
100
+ | dot_spearman | 0.460427 | 0.521260 | +13.21 |
101
+
102
+
103
+ ## Training
104
+ The model was trained with the parameters:
105
+
106
+ **Dataset**
107
+
108
+ We used a collection of datasets of Natural Language Inference as training data:
109
+ - [ESXNLI](https://raw.githubusercontent.com/artetxem/esxnli/master/esxnli.tsv), only the part in spanish
110
+ - [SNLI](https://nlp.stanford.edu/projects/snli/), automatically translated
111
+ - [MultiNLI](https://cims.nyu.edu/~sbowman/multinli/), automatically translated
112
+
113
+ The whole dataset used is available [here](https://huggingface.co/datasets/hackathon-pln-es/nli-es).
114
+
115
+ Here we leave the trick we used to increase the amount of data for training here:
116
+ ```
117
+ for row in reader:
118
+ if row['language'] == 'es':
119
+
120
+ sent1 = row['sentence1'].strip()
121
+ sent2 = row['sentence2'].strip()
122
+
123
+ add_to_samples(sent1, sent2, row['gold_label'])
124
+ add_to_samples(sent2, sent1, row['gold_label']) #Also add the opposite
125
+ ```
126
+
127
+ **DataLoader**:
128
+
129
+ `sentence_transformers.datasets.NoDuplicatesDataLoader.NoDuplicatesDataLoader`
130
+ of length 1818 with parameters:
131
+ ```
132
+ {'batch_size': 64}
133
+ ```
134
+
135
+ **Loss**:
136
+
137
+ `sentence_transformers.losses.MultipleNegativesRankingLoss.MultipleNegativesRankingLoss` with parameters:
138
+ ```
139
+ {'scale': 20.0, 'similarity_fct': 'cos_sim'}
140
+ ```
141
+
142
+ Parameters of the fit()-Method:
143
+ ```
144
+ {
145
+ "epochs": 10,
146
+ "evaluation_steps": 0,
147
+ "evaluator": "sentence_transformers.evaluation.EmbeddingSimilarityEvaluator.EmbeddingSimilarityEvaluator",
148
+ "max_grad_norm": 1,
149
+ "optimizer_class": "<class 'transformers.optimization.AdamW'>",
150
+ "optimizer_params": {
151
+ "lr": 2e-05
152
+ },
153
+ "scheduler": "WarmupLinear",
154
+ "steps_per_epoch": null,
155
+ "warmup_steps": 909,
156
+ "weight_decay": 0.01
157
+ }
158
+ ```
159
+
160
+
161
+ ## Full Model Architecture
162
+ ```
163
+ SentenceTransformer(
164
+ (0): Transformer({'max_seq_length': 514, 'do_lower_case': False}) with Transformer model: RobertaModel
165
+ (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False})
166
+ )
167
+ ```
168
+
169
+ ## Authors
170
+
171
+ [Anibal Pérez](https://huggingface.co/Anarpego),
172
+
173
+ [Emilio Tomás Ariza](https://huggingface.co/medardodt),
174
+
175
+ [Lautaro Gesuelli](https://huggingface.co/Lgesuelli) y
176
+
177
+ [Mauricio Mazuecos](https://huggingface.co/mmazuecos).
config.json ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "./hackathon-pln-es_bertin-roberta-base-finetuning-esnli/",
3
+ "architectures": [
4
+ "RobertaModel"
5
+ ],
6
+ "attention_probs_dropout_prob": 0.1,
7
+ "bos_token_id": 0,
8
+ "classifier_dropout": null,
9
+ "eos_token_id": 2,
10
+ "gradient_checkpointing": false,
11
+ "hidden_act": "gelu",
12
+ "hidden_dropout_prob": 0.1,
13
+ "hidden_size": 768,
14
+ "initializer_range": 0.02,
15
+ "intermediate_size": 3072,
16
+ "layer_norm_eps": 1e-05,
17
+ "max_position_embeddings": 514,
18
+ "model_type": "roberta",
19
+ "num_attention_heads": 12,
20
+ "num_hidden_layers": 12,
21
+ "pad_token_id": 1,
22
+ "position_embedding_type": "absolute",
23
+ "torch_dtype": "float32",
24
+ "transformers_version": "4.19.2",
25
+ "type_vocab_size": 1,
26
+ "use_cache": true,
27
+ "vocab_size": 50265
28
+ }
config_sentence_transformers.json ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
1
+ {
2
+ "__version__": {
3
+ "sentence_transformers": "2.2.0",
4
+ "transformers": "4.17.0",
5
+ "pytorch": "1.10.2"
6
+ }
7
+ }
merges.txt ADDED
The diff for this file is too large to render. See raw diff
modules.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "idx": 0,
4
+ "name": "0",
5
+ "path": "",
6
+ "type": "sentence_transformers.models.Transformer"
7
+ },
8
+ {
9
+ "idx": 1,
10
+ "name": "1",
11
+ "path": "1_Pooling",
12
+ "type": "sentence_transformers.models.Pooling"
13
+ }
14
+ ]
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:56cb66330b3a5d995668381f0b2347cc8ddb67e27c60cefa75f1ce3e6f33b6f8
3
+ size 498649201
sentence_bert_config.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
1
+ {
2
+ "max_seq_length": 514,
3
+ "do_lower_case": false
4
+ }
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
1
+ {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "<unk>", "sep_token": "</s>", "pad_token": "<pad>", "cls_token": "<s>", "mask_token": {"content": "<mask>", "single_word": false, "lstrip": true, "rstrip": false, "normalized": false}}
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
1
+ {"errors": "replace", "bos_token": "<s>", "eos_token": "</s>", "sep_token": "</s>", "cls_token": "<s>", "unk_token": "<unk>", "pad_token": "<pad>", "mask_token": "<mask>", "add_prefix_space": false, "trim_offsets": true, "special_tokens_map_file": null, "name_or_path": "./hackathon-pln-es_bertin-roberta-base-finetuning-esnli/", "tokenizer_class": "RobertaTokenizer"}
vocab.json ADDED
The diff for this file is too large to render. See raw diff