luanafelbarros/TriLingual-BERT-Distil

SentenceTransformer based on google-bert/bert-base-multilingual-cased

This is a sentence-transformers model finetuned from google-bert/bert-base-multilingual-cased. It maps sentences & paragraphs to a 768-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: google-bert/bert-base-multilingual-cased
• Maximum Sequence Length: 128 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity

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': 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, '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("luanafelbarros/TriLingual-BERT-Distil")
# Run inference
sentences = [
    "And in the audio world that's when the microphone gets too close to its sound source, and then it gets in this self-destructive loop that creates a very unpleasant sound.",
    'Y, en el mundo del audio, es cuando el micrófono se acerca demasiado a su fuente de sonido, y entra en este bucle autodestructivo que crea un sonido muy desagradable.',
    'Esta es una mina de Zimbabwe en este momento.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

Evaluation

Metrics

Knowledge Distillation

• Datasets: MSE-val-en-es, MSE-val-en-pt and MSE-val-en-pt-br
• Evaluated with MSEEvaluator

Metric	MSE-val-en-es	MSE-val-en-pt	MSE-val-en-pt-br
negative_mse	-29.5115	-29.9136	-27.7322

Training Details

Training Dataset

Unnamed Dataset

• Size: 3,560,698 training samples
• Columns: english, non_english, and label
• Approximate statistics based on the first 1000 samples:

  	english	non_english	label
  type	string	string	list
  details	min: 4 tokens mean: 25.46 tokens max: 128 tokens	min: 4 tokens mean: 26.67 tokens max: 128 tokens	size: 768 elements

• Samples:

  english	non_english	label
  And then there are certain conceptual things that can also benefit from hand calculating, but I think they're relatively small in number.	Y luego hay ciertas aspectos conceptuales que pueden beneficiarse del cálculo a mano pero creo que son relativamente pocos.	[-0.04180986061692238, 0.12620249390602112, -0.14501447975635529, 0.09695684909820557, -0.10850819200277328, ...]
  One thing I often ask about is ancient Greek and how this relates.	Algo que pregunto a menudo es sobre el griego antiguo y cómo se relaciona.	[0.0034368489868938923, -0.02741478756070137, -0.09426739811897278, 0.04873204976320267, -0.008266829885542393, ...]
  See, the thing we're doing right now is we're forcing people to learn mathematics.	Vean, lo que estamos haciendo ahora es forzar a la gente a aprender matemáticas.	[-0.05048828944563866, 0.2713043689727783, 0.024581076577305794, -0.07316197454929352, -0.044288791716098785, ...]

• Loss: main.ModifiedMatryoshkaLoss with these parameters:

  {
      "loss": "MSELoss",
      "matryoshka_dims": [
          768,
          512,
          256,
          128,
          64
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Evaluation Dataset

Unnamed Dataset

• Size: 6,974 evaluation samples
• Columns: english, non_english, and label
• Approximate statistics based on the first 1000 samples:

  	english	non_english	label
  type	string	string	list
  details	min: 4 tokens mean: 25.68 tokens max: 128 tokens	min: 4 tokens mean: 27.31 tokens max: 128 tokens	size: 768 elements

• Samples:

  english	non_english	label
  Thank you so much, Chris.	Muchas gracias Chris.	[-0.1432434469461441, -0.10335833579301834, -0.07549277693033218, -0.1542435735464096, 0.009247343055903912, ...]
  And it's truly a great honor to have the opportunity to come to this stage twice; I'm extremely grateful.	Y es en verdad un gran honor tener la oportunidad de venir a este escenario por segunda vez. Estoy extremadamente agradecido.	[0.02740730345249176, -0.0601208470761776, -0.023767368867993355, 0.02245006151497364, 0.007412586361169815, ...]
  I have been blown away by this conference, and I want to thank all of you for the many nice comments about what I had to say the other night.	He quedado conmovido por esta conferencia, y deseo agradecer a todos ustedes sus amables comentarios acerca de lo que tenía que decir la otra noche.	[-0.09117366373538971, 0.08627621084451675, -0.05912208557128906, -0.007647979073226452, 0.0008422975661233068, ...]

• Loss: main.ModifiedMatryoshkaLoss with these parameters:

  {
      "loss": "MSELoss",
      "matryoshka_dims": [
          768,
          512,
          256,
          128,
          64
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 200
• per_device_eval_batch_size: 200
• learning_rate: 2e-05
• num_train_epochs: 2
• warmup_ratio: 0.1
• fp16: True
• label_names: ['label']

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 200
• per_device_eval_batch_size: 200
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• 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: 2
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.1
• warmup_steps: 0
• 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
• use_ipex: False
• bf16: False
• fp16: True
• 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: ['label']
• 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}
• deepspeed: None
• label_smoothing_factor: 0.0
• optim: adamw_torch
• optim_args: None
• adafactor: False
• group_by_length: False
• length_column_name: length
• 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: False
• hub_always_push: False
• 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
• dispatch_batches: None
• split_batches: None
• include_tokens_per_second: False
• include_num_input_tokens_seen: False
• neftune_noise_alpha: None
• optim_target_modules: None
• batch_eval_metrics: False
• eval_on_start: False
• use_liger_kernel: False
• eval_use_gather_object: False
• average_tokens_across_devices: False
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	Validation Loss	MSE-val-en-es_negative_mse	MSE-val-en-pt_negative_mse	MSE-val-en-pt-br_negative_mse
0.0562	1000	0.0626	0.0513	-21.2968	-20.7534	-24.2460
0.1123	2000	0.0478	0.0432	-22.1192	-21.8663	-23.2775
0.1685	3000	0.0423	0.0391	-21.6697	-21.5869	-21.6856
0.0562	1000	0.0396	0.0376	-21.7666	-21.7181	-21.6779
0.1123	2000	0.0381	0.0358	-23.4969	-23.5022	-22.9817
0.1685	3000	0.0362	0.0339	-24.7639	-24.8878	-23.8888
0.2247	4000	0.0347	0.0323	-26.5721	-26.7422	-25.4072
0.2808	5000	0.0332	0.0310	-27.6024	-27.8268	-26.4132
0.3370	6000	0.0321	0.0299	-27.7974	-28.0294	-26.6213
0.3932	7000	0.0312	0.0292	-28.2719	-28.4834	-27.0468
0.4493	8000	0.0305	0.0285	-28.2561	-28.5574	-26.8752
0.5055	9000	0.0299	0.0280	-28.6342	-28.9112	-27.2933
0.5617	10000	0.0294	0.0275	-28.5512	-28.8469	-27.1072
0.6178	11000	0.029	0.0271	-28.6788	-28.9608	-27.2056
0.6740	12000	0.0286	0.0267	-29.0159	-29.3281	-27.4770
0.7302	13000	0.0283	0.0264	-28.9224	-29.2461	-27.3500
0.7863	14000	0.028	0.0261	-29.1044	-29.4303	-27.4377
0.8425	15000	0.0277	0.0259	-29.2340	-29.5758	-27.6223
0.8987	16000	0.0275	0.0257	-29.1356	-29.4699	-27.4667
0.9548	17000	0.0273	0.0255	-29.3281	-29.6671	-27.7174
1.0110	18000	0.0271	0.0253	-29.2991	-29.6635	-27.6675
1.0672	19000	0.0268	0.0251	-29.3581	-29.7326	-27.6587
1.1233	20000	0.0266	0.0250	-29.4233	-29.7941	-27.7913
1.1795	21000	0.0265	0.0248	-29.3941	-29.7583	-27.6951
1.2357	22000	0.0264	0.0247	-29.5963	-29.9737	-27.9191
1.2918	23000	0.0262	0.0245	-29.4587	-29.8472	-27.7702
1.3480	24000	0.0262	0.0244	-29.4977	-29.8868	-27.8142
1.4042	25000	0.026	0.0244	-29.5356	-29.9184	-27.8426
1.4603	26000	0.0259	0.0243	-29.5614	-29.9388	-27.8360
1.5165	27000	0.0259	0.0242	-29.5362	-29.9353	-27.8223
1.5727	28000	0.0258	0.0241	-29.5088	-29.9043	-27.7884
1.6288	29000	0.0258	0.0241	-29.4550	-29.8543	-27.6788
1.6850	30000	0.0257	0.0240	-29.5373	-29.9282	-27.7855
1.7412	31000	0.0256	0.0239	-29.5195	-29.9096	-27.7866
1.7973	32000	0.0256	0.0239	-29.5292	-29.9266	-27.7579
1.8535	33000	0.0256	0.0239	-29.5202	-29.9196	-27.7408
1.9097	34000	0.0255	0.0239	-29.5090	-29.9126	-27.7311
1.9659	35000	0.0255	0.0238	-29.5115	-29.9136	-27.7322

Framework Versions

• Python: 3.10.12
• Sentence Transformers: 3.3.1
• Transformers: 4.46.3
• PyTorch: 2.5.1+cu121
• Accelerate: 1.1.1
• Datasets: 3.2.0
• Tokenizers: 0.20.3

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",
}