BumbaLM-Embedding-4B-v0.1 (Jurídico Brasileiro)

Este modelo é uma versão fine-tuned do Qwen/Qwen3-Embedding-4B, especializado no domínio jurídico brasileiro. Ele foi desenvolvido como parte de uma pesquisa de doutorado focada no Enriquecimento de Embeddings Neurais para a Linguagem Jurídica Brasileira.

O BumbaLM-Embedding mapeia sentenças e parágrafos (como peças jurídicas, jurisprudência e consultas processuais) para um espaço vetorial denso de 2560 dimensões, sendo ideal para tarefas de:

• Recuperação de Informação Legal (Legal IR)
• Busca Semântica em processos jurídicos(TJMA e outros tribunais)
• Aplicações de RAG (Retrieval-Augmented Generation) para Direito

Model Description

• Model Type: Sentence Transformer
• Base model: Qwen3-Embedding-4B -->
• Maximum Sequence Length: 1024 tokens
• Output Dimensionality: 2560 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset: Legal Documents
• Language: Portuguese
• License: Unknown

Métricas de Avaliação

O modelo foi avaliado em um conjunto de teste reservado ("bumba_test_eval") composto por pares de consultas e trechos jurídicos reais.

Nota: O modelo base foi refinado utilizando a função de perda TripletLoss, focando na distinção entre passagens juridicamente relevantes e "negativos difíceis" (trechos com termos similares mas semanticamente incorretos para a consulta).

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': 1024, 'do_lower_case': False, 'architecture': 'Qwen3Model'})
  (1): Pooling({'word_embedding_dimension': 2560, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': True, '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("fabricioalmeida/BumbaLM-Embedding-4B-v0.1")
# Run inference
frases = [
    "O réu apresentou habeas corpus preventivo.",
    "A jurisprudência do STJ é pacífica nesse sentido.",
    "Receita de bolo de cenoura com chocolate."
]

embeddings = model.encode(frases)
print(embeddings.shape)
# [3, 2560]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.2075, 0.2365],
#        [0.2075, 1.0000, 0.1745],
#        [0.2365, 0.1745, 1.0000]])

Evaluation

Metrics

Information Retrieval

• Dataset: bumba_test_eval
• Evaluated with InformationRetrievalEvaluator

Training Dataset

Legal Dataset

• Size: 19,500 training samples

• Columns: sentence_0, sentence_1, and sentence_2

• Approximate statistics based on the first 1000 samples:

  	sentence_0	sentence_1	sentence_2
  type	string	string	string
  details	min: 31 tokensmean: 47.57 tokensmax: 92 tokens	min: 6 tokensmean: 242.04 tokensmax: 463 tokens	min: 6 tokensmean: 224.16 tokensmax: 457 tokens

• Loss: TripletLoss with these parameters:

  {
      "distance_metric": "TripletDistanceMetric.EUCLIDEAN",
      "triplet_margin": 5
  }
  

Training Details

Training Hyperparameters

• eval_strategy: steps
• num_train_epochs: 1
• multi_dataset_batch_sampler: round_robin

All Hyperparameters

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 8
• per_device_eval_batch_size: 8
• 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: 5e-05
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1
• num_train_epochs: 1
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.0
• 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
• bf16: False
• 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: batch_sampler
• multi_dataset_batch_sampler: round_robin
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Framework Versions

• Python: 3.12.12
• Sentence Transformers: 5.1.2
• Transformers: 4.57.2
• PyTorch: 2.9.0+cu126
• Accelerate: 1.12.0
• Datasets: 4.4.1
• Tokenizers: 0.22.1

Citation

BibTeX

@misc{bumbalm2025,
    title={BumbaLM-Embeddings: Enriquecimento de Embeddings Neurais para a Linguagem Jurídica Brasileira},
    author={Almeida, Fabrício},
    year={2025},
    description={Modelo de embedding fine-tuned para o domínio jurídico brasileiro (TJMA).}
}