bnkc123/bge-base-financial-matryoshka

BGE base Financial Matryoshka

This is a sentence-transformers model finetuned from BAAI/bge-base-en-v1.5 on the finanical-rag-embedding-dataset dataset. 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: BAAI/bge-base-en-v1.5
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • finanical-rag-embedding-dataset
• Language: en
• License: apache-2.0

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': True}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': True, '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': False, 'include_prompt': True})
  (2): Normalize()
)

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("bnkc123/bge-base-financial-matryoshka")
# Run inference
sentences = [
    "What is the maximum leverage ratio allowed before default under the company's credit facility?",
    "If the company's leverage ratio exceeds 3.50 to 1, it would be in default of its revolving credit facility, impairing its ability to borrow under the facility.",
    'Research and Development Because the industries in which the Company competes are characterized by rapid technological advances, the Company’s ability to compete successfully depends heavily upon its ability to ensure a continual and timely flow of competitive products, services and technologies to the marketplace.',
]
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

Information Retrieval

• Dataset: dim_768
• Evaluated with InformationRetrievalEvaluator with these parameters:

  {
      "truncate_dim": 768
  }
  

Metric	Value
cosine_accuracy@1	0.6771
cosine_accuracy@3	0.8371
cosine_accuracy@5	0.8686
cosine_accuracy@10	0.9186
cosine_precision@1	0.6771
cosine_precision@3	0.279
cosine_precision@5	0.1737
cosine_precision@10	0.0919
cosine_recall@1	0.6771
cosine_recall@3	0.8371
cosine_recall@5	0.8686
cosine_recall@10	0.9186
cosine_ndcg@10	0.8008
cosine_mrr@10	0.7627
cosine_map@100	0.7656

Information Retrieval

• Dataset: dim_512
• Evaluated with InformationRetrievalEvaluator with these parameters:

  {
      "truncate_dim": 512
  }
  

Metric	Value
cosine_accuracy@1	0.6829
cosine_accuracy@3	0.8371
cosine_accuracy@5	0.8757
cosine_accuracy@10	0.92
cosine_precision@1	0.6829
cosine_precision@3	0.279
cosine_precision@5	0.1751
cosine_precision@10	0.092
cosine_recall@1	0.6829
cosine_recall@3	0.8371
cosine_recall@5	0.8757
cosine_recall@10	0.92
cosine_ndcg@10	0.8044
cosine_mrr@10	0.7671
cosine_map@100	0.77

Information Retrieval

• Dataset: dim_256
• Evaluated with InformationRetrievalEvaluator with these parameters:

  {
      "truncate_dim": 256
  }
  

Metric	Value
cosine_accuracy@1	0.6757
cosine_accuracy@3	0.8229
cosine_accuracy@5	0.8643
cosine_accuracy@10	0.9186
cosine_precision@1	0.6757
cosine_precision@3	0.2743
cosine_precision@5	0.1729
cosine_precision@10	0.0919
cosine_recall@1	0.6757
cosine_recall@3	0.8229
cosine_recall@5	0.8643
cosine_recall@10	0.9186
cosine_ndcg@10	0.7984
cosine_mrr@10	0.7599
cosine_map@100	0.7625

Information Retrieval

• Dataset: dim_128
• Evaluated with InformationRetrievalEvaluator with these parameters:

  {
      "truncate_dim": 128
  }
  

Metric	Value
cosine_accuracy@1	0.6714
cosine_accuracy@3	0.8114
cosine_accuracy@5	0.8486
cosine_accuracy@10	0.9014
cosine_precision@1	0.6714
cosine_precision@3	0.2705
cosine_precision@5	0.1697
cosine_precision@10	0.0901
cosine_recall@1	0.6714
cosine_recall@3	0.8114
cosine_recall@5	0.8486
cosine_recall@10	0.9014
cosine_ndcg@10	0.7873
cosine_mrr@10	0.7507
cosine_map@100	0.7543

Information Retrieval

• Dataset: dim_64
• Evaluated with InformationRetrievalEvaluator with these parameters:

  {
      "truncate_dim": 64
  }
  

Metric	Value
cosine_accuracy@1	0.6243
cosine_accuracy@3	0.7843
cosine_accuracy@5	0.82
cosine_accuracy@10	0.8829
cosine_precision@1	0.6243
cosine_precision@3	0.2614
cosine_precision@5	0.164
cosine_precision@10	0.0883
cosine_recall@1	0.6243
cosine_recall@3	0.7843
cosine_recall@5	0.82
cosine_recall@10	0.8829
cosine_ndcg@10	0.7546
cosine_mrr@10	0.7135
cosine_map@100	0.7174

Training Details

Training Dataset

finanical-rag-embedding-dataset

• Dataset: finanical-rag-embedding-dataset at e0b1781
• Size: 6,300 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 7 tokens mean: 20.5 tokens max: 43 tokens	min: 9 tokens mean: 46.09 tokens max: 512 tokens

• Samples:

  anchor	positive
  What was the amount of premiums written by Berkshire Hathaway's Insurance Underwriting in 2023, and how did it compare to the previous year?	Premiums written increased $3.5 billion (24.1%) in 2023 compared to 2022. The increase was primarily due to RSUI and CapSpecialty ($2.1 billion), as well as comparative increases from BHSI and BH Direct, and to a lesser extent the other businesses. Premiums written
  What types of transportation equipment does XTRA Corporation manage in its fleet?	XTRA manages a diverse fleet of approximately 90,000 units located at 47 facilities throughout the U.S. The fleet includes over-the-road and storage trailers, chassis, temperature-controlled vans and flatbed trailers.
  What seasonal trends affect the company's sales volumes?	Sales volumes for the company are highest in the second fiscal quarter due to seasonal influences, particularly during the spring season in the regions it serves.

• Loss: MatryoshkaLoss with these parameters:

  {
      "loss": "MultipleNegativesRankingLoss",
      "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: epoch
• per_device_train_batch_size: 32
• per_device_eval_batch_size: 16
• gradient_accumulation_steps: 16
• learning_rate: 2e-05
• num_train_epochs: 4
• lr_scheduler_type: cosine
• warmup_ratio: 0.1
• bf16: True
• tf32: True
• load_best_model_at_end: True
• optim: adamw_torch_fused
• push_to_hub: True
• hub_model_id: bnkc123/bge-base-financial-matryoshka
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: epoch
• prediction_loss_only: True
• per_device_train_batch_size: 32
• per_device_eval_batch_size: 16
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 16
• 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: 4
• max_steps: -1
• lr_scheduler_type: cosine
• 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: True
• fp16: False
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: True
• 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: True
• 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}
• tp_size: 0
• 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_fused
• 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: True
• resume_from_checkpoint: None
• hub_model_id: bnkc123/bge-base-financial-matryoshka
• hub_strategy: every_save
• hub_private_repo: None
• 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
• 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: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	dim_768_cosine_ndcg@10	dim_512_cosine_ndcg@10	dim_256_cosine_ndcg@10	dim_128_cosine_ndcg@10	dim_64_cosine_ndcg@10
0.8122	10	25.483	-	-	-	-	-
1.0	13	-	0.7890	0.7887	0.7815	0.7647	0.7280
1.5685	20	9.1323	-	-	-	-	-
2.0	26	-	0.7952	0.7982	0.7933	0.7801	0.7477
2.3249	30	6.7535	-	-	-	-	-
3.0	39	-	0.8019	0.8048	0.7989	0.7865	0.7547
3.0812	40	6.5646	-	-	-	-	-
3.731	48	-	0.8008	0.8044	0.7984	0.7873	0.7546

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.12.6
• Sentence Transformers: 4.1.0
• Transformers: 4.51.3
• PyTorch: 2.7.0+cu126
• Accelerate: 1.6.0
• Datasets: 3.5.1
• Tokenizers: 0.21.1

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",
}

MatryoshkaLoss

@misc{kusupati2024matryoshka,
    title={Matryoshka Representation Learning},
    author={Aditya Kusupati and Gantavya Bhatt and Aniket Rege and Matthew Wallingford and Aditya Sinha and Vivek Ramanujan and William Howard-Snyder and Kaifeng Chen and Sham Kakade and Prateek Jain and Ali Farhadi},
    year={2024},
    eprint={2205.13147},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

MultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}