jzju/sbert-sv-lim2

Usar este modelo es fácil si tienes sentence-transformers instalado:

pip install -U sentence-transformers

Luego puedes usar el modelo así:

from sentence_transformers import SentenceTransformer
sentences = ['Esta es una oración de ejemplo', 'Cada oración se convierte']
model = SentenceTransformer('jzju/sbert-sv-lim2')
embeddings = model.encode(sentences)
print(embeddings)

Código de entrenamiento:

from datasets import load_dataset, concatenate_datasets
from sentence_transformers import (SentenceTransformer, InputExample, losses, models, util, datasets)
from torch.utils.data import DataLoader
from torch import nn
import random
word_embedding_model = models.Transformer('KBLab/bert-base-swedish-cased-new', max_seq_length=256)
pooling_model = models.Pooling(word_embedding_model.get_word_embedding_dimension())
dense_model = models.Dense(in_features=pooling_model.get_sentence_embedding_dimension(), out_features=256, activation_function=nn.Tanh())
model = SentenceTransformer(modules=[word_embedding_model, pooling_model, dense_model])
def pair():
    def norm(x):
        x['label'] = x['label'] / m
        return x
    dd = []
    for sub in ['swepar', 'swesim_relatedness', 'swesim_similarity']:
        ds = concatenate_datasets([d for d in load_dataset('sbx/superlim-2', sub).values()])
        if 'sentence_1' in ds.features:
            ds = ds.rename_column('sentence_1', 'd1')
            ds = ds.rename_column('sentence_2', 'd2')
        else:
            ds = ds.rename_column('word_1', 'd1')
            ds = ds.rename_column('word_2', 'd2')
        m = max([d['label'] for d in ds])
        dd.append(ds.map(norm))
    ds = concatenate_datasets(dd)
    train_examples = []
    for d in ds:
        train_examples.append(InputExample(texts=[d['d1'], d['d2']], label=d['label']))
    train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=64)
    train_loss = losses.CosineSimilarityLoss(model)
    model.fit(train_objectives=[(train_dataloader, train_loss)], epochs=10, warmup_steps=100)

def nli():
    ds = concatenate_datasets([d for d in load_dataset('sbx/superlim-2', 'swenli').values()])
    def add_to_samples(sent1, sent2, label):
        if sent1 not in train_data:
            train_data[sent1] = {0: set(), 1: set(), 2: set()}
        train_data[sent1][label].add(sent2)
    train_data = {}
    for d in ds:
        add_to_samples(d['premise'], d['hypothesis'], d['label'])
        add_to_samples(d['hypothesis'], d['premise'], d['label'])
    train_samples = []
    for sent1, others in train_data.items():
        if len(others[0]) > 0 and len(others[1]) > 0:
            train_samples.append(InputExample(texts=[sent1, random.choice(list(others[0])), random.choice(list(others[1]))]))
        train_samples.append(InputExample(texts=[random.choice(list(others[0])), sent1, random.choice(list(others[1]))]))
    train_dataloader = datasets.NoDuplicatesDataLoader(train_samples, batch_size=64)
    train_loss = losses.MultipleNegativesRankingLoss(model)
    model.fit(train_objectives=[(train_dataloader, train_loss)], epochs=1, warmup_steps=100)
pair()
nli()
model.save()