Moco Sentencedistilbertv2.0

🚀 moco-sentencedistilbertV2.0

This is a model that maps sentences & paragraphs to a 768-dimensional dense vector space, suitable for tasks like clustering or semantic search.

Pipeline and Tags

• Pipeline Tag: sentence-similarity
• Tags: sentence-transformers, feature-extraction, sentence-similarity, transformers, ko, en

Widget Example

• Source Sentence: "대한민국의 수도는?"
• Comparison Sentences:
  • "서울특별시는 한국이 정치,경제,문화 중심 도시이다."
  • "부산은 대한민국의 제2의 도시이자 최대의 해양 물류 도시이다."
  • "제주도는 대한민국에서 유명한 관광지이다"
  • "Seoul is the capital of Korea"
  • "울산광역시는 대한민국 남동부 해안에 있는 광역시이다"

🚀 Quick Start

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

• This model is created by first training the mdistilbertV1.1 model with the moco-corpus (3.2M sentences extracted by MOCOMSYS) using SentenceBERT, and then performing additional STS teacher-student distillation training.
• Vocab: 164,314 (17,870 new vocab added to the original mdistilbertV1.1 vocab of 146,444).
• MLM Model: bongsoo/mdistilbertV2.0

📦 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-sentencedistilbertV2.0')
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

[[ 9.7172342e-02 -3.3226651e-01 -7.7130608e-05 ...  1.3900512e-02 2.1072578e-01 -1.5386048e-01]
 [ 2.3313640e-02 -8.4675789e-02 -3.7715461e-06 ...  2.4005771e-02 -1.6602692e-01 -1.2729791e-01]]
*cosine_score:0.3383665680885315

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-sentencedistilbertV2.0')
model = AutoModel.from_pretrained('bongsoo/moco-sentencedistilbertV2.0')

# 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([[ 9.7172e-02, -3.3227e-01, -7.7131e-05,  ...,  1.3901e-02, 2.1073e-01, -1.5386e-01],
        [ 2.3314e-02, -8.4676e-02, -3.7715e-06,  ...,  2.4006e-02, -1.6603e-01, -1.2730e-01]])
*cosine_score:0.3383665680885315

📚 Documentation

Evaluation Results

• The corpora used for performance measurement are the following Korean (kor) and English (en) evaluation corpora:
  • Korean: korsts (1,379 sentence pairs) and klue-sts (519 sentence pairs)
  • English: stsb_multi_mt (1,376 sentence pairs)
• The performance metric is cosin.spearman for comparison.
• Refer to the evaluation 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: bongsoo/mdistilbertV1.1 (distilbert-base-multilingual-cased trained on the kowiki20220620 (4.4M) corpus)
• Corpus: nlp_corpus (3.2M): A refined corpus from MOCOMSYS files
• Hyperparameters: Learning Rate: 5e-5, Epochs: 8, Batch Size: 32, Max Token Length: 128
• Output Model: mdistilbertV2.0
• Training Time: 27h
• Refer to the training code here

2. STS Training

• Convert DistilBERT to SentenceBERT.
• Input Model: mdistilbertV2.0
• Corpus: korsts + kluestsV1.1 + stsb_multi_mt + mteb/sickr-sts (Total: 33,093)
• Hyperparameters: Learning Rate: 2e-5, Epochs: 200, Batch Size: 32, Max Token Length: 128
• Output Model: sbert-mdistilbertV2.0
• Training Time: 5h
• Refer to the training code here

3. Distillation Training

• Student Model: sbert-mdistilbertV2.0
• Teacher Model: paraphrase-multilingual-mpnet-base-v2
• Corpus: en_ko_train.tsv (A parallel corpus of Korean-English social science fields: 1.1M)
• Hyperparameters: Learning Rate: 5e-5, Epochs: 40, Batch Size: 32, Max Token Length: 128
• Output Model: sbert-mdistilbertV2.0.2-distil
• Training Time: 11h
• Refer to the training code here

4. STS Training

• Train the SentenceBERT model on STS.
• Input Model: sbert-mdistilbertV2.0.2-distil
• Corpus: korsts + kluestsV1.1 + stsb_multi_mt + mteb/sickr-sts (Total: 33,093)
• Hyperparameters: Learning Rate: 3e-5, Epochs: 800, Batch Size: 32, Max Token Length: 128
• Output Model: moco-sentencedistilbertV2.0
• Training Time: 15h
• Refer to the training code here

For more details about the model creation process, refer to 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-mdistilbertV2.0.2-distil",
  "activation": "gelu",
  "architectures": [
    "DistilBertModel"
  ],
  "attention_dropout": 0.1,
  "dim": 768,
  "dropout": 0.1,
  "hidden_dim": 3072,
  "initializer_range": 0.02,
  "max_position_embeddings": 512,
  "model_type": "distilbert",
  "n_heads": 12,
  "n_layers": 6,
  "output_past": true,
  "pad_token_id": 0,
  "qa_dropout": 0.1,
  "seq_classif_dropout": 0.2,
  "sinusoidal_pos_embds": false,
  "tie_weights_": true,
  "torch_dtype": "float32",
  "transformers_version": "4.21.2",
  "vocab_size": 164314
}

🔧 Technical Details

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

This model is developed by bongsoo.