philschmid/bge-base-financial-matryoshka

BGE base Financial Matryoshka

This is a sentence-transformers model finetuned from BAAI/bge-base-en-v1.5. 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 tokens
• Similarity Function: Cosine Similarity
• 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("philschmid/bge-base-financial-matryoshka")
# Run inference
sentences = [
    "What was Gilead's total revenue in 2023?",
    'What was the total revenue for the year ended December 31, 2023?',
    'How much was the impairment related to the CAT loan receivable in 2023?',
]
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: basline_768
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.7086
cosine_accuracy@3	0.8514
cosine_accuracy@5	0.8843
cosine_accuracy@10	0.9271
cosine_precision@1	0.7086
cosine_precision@3	0.2838
cosine_precision@5	0.1769
cosine_precision@10	0.0927
cosine_recall@1	0.7086
cosine_recall@3	0.8514
cosine_recall@5	0.8843
cosine_recall@10	0.9271
cosine_ndcg@10	0.8215
cosine_mrr@10	0.7874
cosine_map@100	0.7907

Information Retrieval

• Dataset: basline_512
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.7114
cosine_accuracy@3	0.85
cosine_accuracy@5	0.8829
cosine_accuracy@10	0.9229
cosine_precision@1	0.7114
cosine_precision@3	0.2833
cosine_precision@5	0.1766
cosine_precision@10	0.0923
cosine_recall@1	0.7114
cosine_recall@3	0.85
cosine_recall@5	0.8829
cosine_recall@10	0.9229
cosine_ndcg@10	0.8209
cosine_mrr@10	0.7879
cosine_map@100	0.7916

Information Retrieval

• Dataset: basline_256
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.7057
cosine_accuracy@3	0.8414
cosine_accuracy@5	0.88
cosine_accuracy@10	0.9229
cosine_precision@1	0.7057
cosine_precision@3	0.2805
cosine_precision@5	0.176
cosine_precision@10	0.0923
cosine_recall@1	0.7057
cosine_recall@3	0.8414
cosine_recall@5	0.88
cosine_recall@10	0.9229
cosine_ndcg@10	0.8162
cosine_mrr@10	0.7818
cosine_map@100	0.7854

Information Retrieval

• Dataset: basline_128
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.7029
cosine_accuracy@3	0.8343
cosine_accuracy@5	0.8743
cosine_accuracy@10	0.9171
cosine_precision@1	0.7029
cosine_precision@3	0.2781
cosine_precision@5	0.1749
cosine_precision@10	0.0917
cosine_recall@1	0.7029
cosine_recall@3	0.8343
cosine_recall@5	0.8743
cosine_recall@10	0.9171
cosine_ndcg@10	0.8109
cosine_mrr@10	0.7769
cosine_map@100	0.7803

Information Retrieval

• Dataset: basline_64
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6729
cosine_accuracy@3	0.8171
cosine_accuracy@5	0.8614
cosine_accuracy@10	0.9014
cosine_precision@1	0.6729
cosine_precision@3	0.2724
cosine_precision@5	0.1723
cosine_precision@10	0.0901
cosine_recall@1	0.6729
cosine_recall@3	0.8171
cosine_recall@5	0.8614
cosine_recall@10	0.9014
cosine_ndcg@10	0.79
cosine_mrr@10	0.754
cosine_map@100	0.7582

Training Details

Training Dataset

Unnamed Dataset

• Size: 6,300 training samples
• Columns: positive and anchor
• Approximate statistics based on the first 1000 samples:

  	positive	anchor
  type	string	string
  details	min: 10 tokens mean: 46.11 tokens max: 289 tokens	min: 7 tokens mean: 20.26 tokens max: 43 tokens

• Samples:

  positive	anchor
  Fiscal 2023 total gross profit margin of 35.1% represents an increase of 1.7 percentage points as compared to the respective prior year period.	What was the total gross profit margin for Hewlett Packard Enterprise in fiscal 2023?
  Noninterest expense increased to $65.8 billion in 2023, primarily due to higher investments in people and technology and higher FDIC expense, including $2.1 billion for the estimated special assessment amount arising from the closure of Silicon Valley Bank and Signature Bank.	What was the total noninterest expense for the company in 2023?
  As of May 31, 2022, FedEx Office had approximately 12,000 employees.	How many employees did FedEx Office have as of May 31, 2023?

• 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
• 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
• 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}
• 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: 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
• 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
• sanity_evaluation: False
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	basline_128_cosine_map@100	basline_256_cosine_map@100	basline_512_cosine_map@100	basline_64_cosine_map@100	basline_768_cosine_map@100
0.8122	10	1.5259	-	-	-	-	-
0.9746	12	-	0.7502	0.7737	0.7827	0.7185	0.7806
1.6244	20	0.6545	-	-	-	-	-
1.9492	24	-	0.7689	0.7844	0.7869	0.7447	0.7909
2.4365	30	0.4784	-	-	-	-	-
2.9239	36	-	0.7733	0.7916	0.7904	0.7491	0.7930
3.2487	40	0.3827	-	-	-	-	-
3.8985	48	-	0.7739	0.7907	0.7900	0.7479	0.7948
0.8122	10	0.2685	-	-	-	-	-
0.9746	12	-	0.7779	0.7932	0.7948	0.7517	0.7943
1.6244	20	0.183	-	-	-	-	-
1.9492	24	-	0.7784	0.7929	0.7963	0.7575	0.7957
2.4365	30	0.1877	-	-	-	-	-
2.9239	36	-	0.7814	0.7914	0.7992	0.7570	0.7974
3.2487	40	0.1826	-	-	-	-	-
3.8985	48	-	0.7818	0.7916	0.7976	0.7580	0.7960
0.8122	10	0.071	-	-	-	-	-
0.9746	12	-	0.7810	0.7935	0.7954	0.7550	0.7949
1.6244	20	0.0629	-	-	-	-	-
1.9492	24	-	0.7855	0.7914	0.7989	0.7559	0.7981
2.4365	30	0.0827	-	-	-	-	-
2.9239	36	-	0.7893	0.7927	0.7987	0.7539	0.7961
3.2487	40	0.1003	-	-	-	-	-
3.8985	48	-	0.7903	0.7915	0.7980	0.7530	0.7951
0.8122	10	0.0213	-	-	-	-	-
0.9746	12	-	0.7786	0.7869	0.7885	0.7566	0.7908
1.6244	20	0.0234	-	-	-	-	-
1.9492	24	-	0.783	0.7882	0.793	0.7551	0.7946
2.4365	30	0.0357	-	-	-	-	-
2.9239	36	-	0.7838	0.7892	0.7922	0.7579	0.7907
3.2487	40	0.0563	-	-	-	-	-
3.8985	48	-	0.7846	0.7887	0.7912	0.7582	0.7901
0.8122	10	0.0075	-	-	-	-	-
0.9746	12	-	0.7730	0.7816	0.7818	0.7550	0.7868
1.6244	20	0.01	-	-	-	-	-
1.9492	24	-	0.7827	0.785	0.7896	0.7551	0.7915
2.4365	30	0.0154	-	-	-	-	-
2.9239	36	-	0.7808	0.7838	0.7921	0.7584	0.7916
3.2487	40	0.0312	-	-	-	-	-
3.8985	48	-	0.7803	0.7854	0.7916	0.7582	0.7907

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.10.13
• Sentence Transformers: 3.0.0
• Transformers: 4.42.0.dev0
• PyTorch: 2.1.2+cu121
• Accelerate: 0.29.2
• Datasets: 2.19.1
• Tokenizers: 0.19.1

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}
}