SentenceTransformer based on NbAiLab/nb-bert-large

This is a sentence-transformers model finetuned from NbAiLab/nb-bert-large. It builds on the previous work of the existing NbAiLab/nb-sbert-base model, using a larger foundational model and providing a larger max sequence length for inputs.

The model maps sentences & paragraphs to a 1024-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more. The easiest way is to simply measure the cosine distance between two sentences. Sentences that are close to each other in meaning, will have a small cosine distance and a similarity close to 1. The model is trained in such a way that similar sentences in different languages should also be close to each other. Ideally, an English-Norwegian sentence pair should have high similarity.

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: NbAiLab/nb-bert-large
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 1024 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset: Subset of NbAiLab/mnli-norwegian
• Language: Norwegian and English
• License: Apache 2.0

EU AI Act

This release is a non-generative encoder model whose outputs are vectors/scores rather than language or media. Its intended functionality is limited to representation, retrieval, ranking, or classification support. On that basis, the release is preliminarily assessed as not falling within the provider obligations for GPAI models under the EU AI Act definitions, subject to legal confirmation if capability scope or marketed generality changes. For more information, see the Model Documentation Form here.

Model Sources

• Documentation: Sentence Transformers Documentation
• Repository: Sentence Transformers on GitHub
• Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False, 'architecture': 'BertModel'})
  (1): Pooling({'word_embedding_dimension': 1024, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("NbAiLab/nb-sbert-v2-large")
# Run inference
sentences = [
    "This is a Norwegian boy", 
    "Dette er en norsk gutt"
    ]

embeddings = model.encode(sentences)
print(embeddings.shape)
# (2, 1024)

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.9288],
#         [0.9288, 1.0000]])

Direct Usage (Transformers)

Without sentence-transformers, you can still use the model. First, you pass in your input through the transformer model, then you have to apply the right pooling-operation on top of the contextualized word embeddings.

Click to see the direct usage in Transformers

import torch

from sklearn.metrics.pairwise import cosine_similarity
from transformers import AutoTokenizer, AutoModel

#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 a Norwegian boy", "Dette er en norsk gutt"]

# Load model from HuggingFace Hub
tokenizer = AutoTokenizer.from_pretrained('NbAiLab/nb-sbert-v2-large')
model = AutoModel.from_pretrained('NbAiLab/nb-sbert-v2-large')

# 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.
embeddings = mean_pooling(model_output, encoded_input['attention_mask'])
print(embeddings.shape)
# torch.Size([2, 1024])

similarity = cosine_similarity(embeddings[0].reshape(1, -1), embeddings[1].reshape(1, -1))
print(similarity)
# This should give 0.9288 in the example above.

Evaluation

Metrics

Semantic Similarity

• Dataset: STS Benchmark
• Evaluated with EmbeddingSimilarityEvaluator

MTEB (Scandinavian)

Training Details

Training Dataset

Unnamed Dataset

• Size: 527,098 training samples

• Columns: anchor, positive, and negative

• Approximate statistics based on the first 1000 samples:

  	anchor	positive	negative
  type	string	string	string
  details	min: 5 tokens mean: 19.56 tokens max: 128 tokens	min: 5 tokens mean: 19.53 tokens max: 128 tokens	min: 6 tokens mean: 13.39 tokens max: 38 tokens

• Samples:

  anchor	positive	negative
  Det som følger er mindre en glid nedover en glatt skråning enn et profesjonelt skred som resulterer i enten en oppsigelse eller en smal flukt til neste drømmejobb, der, selvfølgelig, syklusen gjentas igjen.	Syklusen gjentar seg ved neste jobb.	Syklusen gjentar seg sjelden ved neste jobb.
  Syklusen gjentar seg ved neste jobb.	Det som følger er mindre en glid nedover en glatt skråning enn et profesjonelt skred som resulterer i enten en oppsigelse eller en smal flukt til neste drømmejobb, der, selvfølgelig, syklusen gjentas igjen.	Syklusen gjentar seg sjelden ved neste jobb.
  The public areas are spectacular, the rooms a bit less so, but a long-awaited renovation was carried out in 1998.	Rommene er fine, men det offentlige området er i en egen liga.	The public area was fine, but the rooms were really something else.
  Ah, but he had no opportunity.	Han hadde ikke sjansen til å gjøre noe.	Han hadde mange muligheter.

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim",
      "gather_across_devices": false,
      "directions": [
          "query_to_doc"
      ],
      "partition_mode": "joint",
      "hardness_mode": null,
      "hardness_strength": 0.0
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• per_device_train_batch_size: 32
• gradient_accumulation_steps: 4
• num_train_epochs: 1
• learning_rate: 2e-05
• warmup_steps: 412
• bf16: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 32
• per_device_eval_batch_size: 32
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 4
• eval_accumulation_steps: None
• torch_empty_cache_steps: None
• learning_rate: 2e-05
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 1
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.0
• warmup_steps: 412
• log_level: passive
• log_level_replica: warning
• log_on_each_node: True
• logging_nan_inf_filter: True
• save_safetensors: True
• save_on_each_node: False
• save_only_model: False
• restore_callback_states_from_checkpoint: False
• no_cuda: False
• use_cpu: False
• use_mps_device: False
• seed: 42
• data_seed: None
• jit_mode_eval: False
• bf16: True
• fp16: False
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• local_rank: 0
• ddp_backend: None
• tpu_num_cores: None
• tpu_metrics_debug: False
• debug: []
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_prefetch_factor: None
• past_index: -1
• disable_tqdm: False
• remove_unused_columns: True
• label_names: None
• load_best_model_at_end: False
• ignore_data_skip: False
• fsdp: []
• fsdp_min_num_params: 0
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• fsdp_transformer_layer_cls_to_wrap: None
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• parallelism_config: None
• deepspeed: None
• label_smoothing_factor: 0.0
• optim: adamw_torch_fused
• optim_args: None
• adafactor: False
• group_by_length: False
• length_column_name: length
• project: huggingface
• trackio_space_id: trackio
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• skip_memory_metrics: True
• use_legacy_prediction_loop: False
• push_to_hub: False
• resume_from_checkpoint: None
• hub_model_id: None
• hub_strategy: every_save
• hub_private_repo: None
• hub_always_push: False
• hub_revision: None
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• include_for_metrics: []
• eval_do_concat_batches: True
• fp16_backend: auto
• push_to_hub_model_id: None
• push_to_hub_organization: None
• mp_parameters:
• auto_find_batch_size: False
• full_determinism: False
• torchdynamo: None
• ray_scope: last
• ddp_timeout: 1800
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• include_tokens_per_second: False
• include_num_input_tokens_seen: no
• neftune_noise_alpha: None
• optim_target_modules: None
• batch_eval_metrics: False
• eval_on_start: False
• use_liger_kernel: False
• liger_kernel_config: None
• eval_use_gather_object: False
• average_tokens_across_devices: True
• prompts: None
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Framework Versions

• Python: 3.12.3
• Sentence Transformers: 5.3.0
• Transformers: 4.57.3
• PyTorch: 2.9.1+rocm6.4
• Accelerate: 1.12.0
• Datasets: 4.6.0
• Tokenizers: 0.22.2

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

MultipleNegativesRankingLoss

@misc{oord2019representationlearningcontrastivepredictive,
      title={Representation Learning with Contrastive Predictive Coding},
      author={Aaron van den Oord and Yazhe Li and Oriol Vinyals},
      year={2019},
      eprint={1807.03748},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/1807.03748},
}

NbAiLab/nb-bert-large

@inproceedings{kummervold-etal-2021-operationalizing,
  title     = {Operationalizing a National Digital Library: The Case for a {N}orwegian Transformer Model},
  author    = {Kummervold, Per E  and
               De la Rosa, Javier  and
               Wetjen, Freddy  and
               Brygfjeld, Svein Arne},
  booktitle = {Proceedings of the 23rd Nordic Conference on Computational Linguistics (NoDaLiDa)},
  year      = {2021},
  address   = {Reykjavik, Iceland (Online)},
  publisher = {Linköping University Electronic Press, Sweden},
  url       = {https://huggingface.co/papers/2104.09617},
  pages     = {20--29},
  abstract  = {In this work, we show the process of building a large-scale training set from digital and digitized collections at a national library.
  The resulting Bidirectional Encoder Representations from Transformers (BERT)-based language model for Norwegian outperforms multilingual BERT (mBERT) models
  in several token and sequence classification tasks for both Norwegian Bokmål and Norwegian Nynorsk. Our model also improves the mBERT performance for other
  languages present in the corpus such as English, Swedish, and Danish. For languages not included in the corpus, the weights degrade moderately while keeping strong multilingual properties. Therefore,
  we show that building high-quality models within a memory institution using somewhat noisy optical character recognition (OCR) content is feasible, and we hope to pave the way for other memory institutions to follow.},
}

Citing & Authors

The model was trained by Victoria Handford and Lucas Georges Gabriel Charpentier. The documentation was initially autogenerated by the SentenceTransformers library then revised by Victoria Handford, Lucas Georges Gabriel Charpentier, and Javier de la Rosa.