Moco Sentencebertv2.0

🚀 moco-sentencebertV2.0

This is a sentence-transformers model that maps sentences and paragraphs to a 768-dimensional dense vector space. It can be used for tasks such as clustering or semantic search.

✨ Features

• This model is created by converting the bongsoo/mbertV2.0 MLM model into SentenceBERT and then performing additional STS teacher-student distillation training.
• Vocab: 152,537 (32,989 new vocabularies are added to the original 119,548 vocabularies).

📦 Installation

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence_transformers

💻 Usage Examples

Basic Usage

from sentence_transformers import SentenceTransformer
sentences = ["This is an example sentence", "Each sentence is converted"]

model = SentenceTransformer('bongsoo/moco-sentencebertV2.0')
embeddings = model.encode(sentences)
print(embeddings)

# Use sklearn to calculate cosine_scores
# => The input embeddings should be 2D, like (1,768).
from sklearn.metrics.pairwise import paired_cosine_distances, paired_euclidean_distances, paired_manhattan_distances
cosine_scores = 1 - (paired_cosine_distances(embeddings[0].reshape(1,-1), embeddings[1].reshape(1,-1)))

print(f'*cosine_score:{cosine_scores[0]}')

Outputs

[[ 0.16649279 -0.2933038  -0.00391259 ...  0.00720964  0.18175027  -0.21052675]
 [ 0.10106096 -0.11454111 -0.00378215 ... -0.009032   -0.2111504   -0.15030429]]
*cosine_score:0.3352515697479248

Advanced Usage

from transformers import AutoTokenizer, AutoModel
import torch


#Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)


# Sentences we want sentence embeddings for
sentences = ['This is an example sentence', 'Each sentence is converted']

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('bongsoo/moco-sentencebertV2.0')
model = AutoModel.from_pretrained('bongsoo/moco-sentencebertV2.0')

# Tokenize sentences
encoded_input = tokenizer(sentences, padding=True, truncation=True, return_tensors='pt')

# Compute token embeddings
with torch.no_grad():
    model_output = model(**encoded_input)

# Perform pooling. In this case, mean pooling.
sentence_embeddings = mean_pooling(model_output, encoded_input['attention_mask'])

print("Sentence embeddings:")
print(sentence_embeddings)

# Use sklearn to calculate cosine_scores
# => The input embeddings should be 2D, like (1,768).
from sklearn.metrics.pairwise import paired_cosine_distances, paired_euclidean_distances, paired_manhattan_distances
cosine_scores = 1 - (paired_cosine_distances(sentence_embeddings[0].reshape(1,-1), sentence_embeddings[1].reshape(1,-1)))

print(f'*cosine_score:{cosine_scores[0]}')

Outputs

Sentence embeddings:
tensor([[ 0.1665, -0.2933, -0.0039,  ...,  0.0072,  0.1818, -0.2105],
        [ 0.1011, -0.1145, -0.0038,  ..., -0.0090, -0.2112, -0.1503]])
*cosine_score:0.3352515697479248

📚 Documentation

Evaluation Results

• For performance measurement, the following Korean (kor) and English (en) evaluation corpora are used:
  • Korean: korsts (1,379 sentence pairs) and klue-sts (519 sentence pairs).
  • English: stsb_multi_mt (1,376 sentence pairs) and glue:stsb (1,500 sentence pairs).
• The performance metric is cosin.spearman for comparison.
• Refer to the evaluation measurement code here.

For an automated evaluation of this model, see the Sentence Embeddings Benchmark: https://seb.sbert.net

Training

The model was trained with the following parameters:

1. MLM Training

• Input model: bert-base-multilingual-cased
• Corpus: Training - bongsoo/moco-corpus-kowiki2022 (7.6M), Evaluation - bongsoo/bongevalsmall
• Hyperparameters: LearningRate - 5e-5, epochs - 8, batchsize - 32, max_token_len - 128
• Output model: mbertV2.0 (size: 813MB)
• Training time: 90h/1GPU (24GB/19.6GB use)
• Loss: Training loss - 2.258400, Evaluation loss - 3.102096, Perplexity - 19.78158 (bong_eval:1,500)
• Refer to the training code here.

2. STS Training => Convert BERT to SentenceBERT.

• Input model: mbertV2.0
• Corpus: korsts + kluestsV1.1 + stsb_multi_mt + mteb/sickr-sts (total: 33,093)
• Hyperparameters: LearningRate - 3e-5, epochs - 200, batchsize - 32, max_token_len - 128
• Output model: sbert-mbertV2.0 (size: 813MB)
• Training time: 9h20m/1GPU (24GB/9.0GB use)
• Loss (cosin_spearman): 0.799 (corpus: korsts(tune_test.tsv))
• Refer to the training code here.

3. Distillation Training

• Student model: sbert-mbertV2.0
• Teacher model: paraphrase-multilingual-mpnet-base-v2
• Corpus: en_ko_train.tsv (Korean-English social science parallel corpus: 1.1M)
• Hyperparameters: LearningRate - 5e-5, epochs - 40, batchsize - 128, max_token_len - 128
• Output model: sbert-mlbertV2.0-distil
• Training time: 17h/1GPU (24GB/18.6GB use)
• Refer to the training code here.

4. STS Training => Train the SentenceBERT model with STS.

• Input model: sbert-mlbertV2.0-distil
• Corpus: korsts (5,749) + kluestsV1.1 (11,668) + stsb_multi_mt (5,749) + mteb/sickr-sts (9,927) + glue stsb (5,749) (total: 38,842)
• Hyperparameters: LearningRate - 3e-5, epochs - 800, batchsize - 64, max_token_len - 128
• Output model: moco-sentencebertV2.0
• Training time: 25h/1GPU (24GB/13GB use)
• Refer to the training code here.

Refer to the detailed content of the model production process here.

DataLoader: torch.utils.data.dataloader.DataLoader of length 1035 with parameters:

{'batch_size': 32, 'sampler': 'torch.utils.data.sampler.RandomSampler', 'batch_sampler': 'torch.utils.data.sampler.BatchSampler'}

Config:

{
  "_name_or_path": "../../data11/model/sbert/sbert-mbertV2.0-distil",
  "architectures": [
    "BertModel"
  ],
  "attention_probs_dropout_prob": 0.1,
  "classifier_dropout": null,
  "directionality": "bidi",
  "hidden_act": "gelu",
  "hidden_dropout_prob": 0.1,
  "hidden_size": 768,
  "initializer_range": 0.02,
  "intermediate_size": 3072,
  "layer_norm_eps": 1e-12,
  "max_position_embeddings": 512,
  "model_type": "bert",
  "num_attention_heads": 12,
  "num_hidden_layers": 12,
  "pad_token_id": 0,
  "pooler_fc_size": 768,
  "pooler_num_attention_heads": 12,
  "pooler_num_fc_layers": 3,
  "pooler_size_per_head": 128,
  "pooler_type": "first_token_transform",
  "position_embedding_type": "absolute",
  "torch_dtype": "float32",
  "transformers_version": "4.21.2",
  "type_vocab_size": 2,
  "use_cache": true,
  "vocab_size": 152537
}

🔧 Technical Details

SentenceTransformer(
  (0): Transformer({'max_seq_length': 128, 'do_lower_case': False}) with Transformer model: BertModel 
  (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})
)

📄 License

No license information provided in the original document.

Citing & Authors

bongsoo