anikulkar/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("anikulkar/bge-base-financial-matryoshka")
# Run inference
sentences = [
    'We recognize gains and losses on pension and postretirement plan assets and obligations immediately in Other income (expense) - net in our consolidated statements of income.',
    'Where are gains and losses on pension and postretirement plan assets and obligations recognized in financial statements?',
    'What is the total amount of property, plant, and equipment, net, reported by the company for the fiscal year 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: dim_768
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6829
cosine_accuracy@3	0.8229
cosine_accuracy@5	0.86
cosine_accuracy@10	0.9057
cosine_precision@1	0.6829
cosine_precision@3	0.2743
cosine_precision@5	0.172
cosine_precision@10	0.0906
cosine_recall@1	0.6829
cosine_recall@3	0.8229
cosine_recall@5	0.86
cosine_recall@10	0.9057
cosine_ndcg@10	0.7961
cosine_mrr@10	0.7608
cosine_map@100	0.7647

Information Retrieval

• Dataset: dim_512
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6843
cosine_accuracy@3	0.8229
cosine_accuracy@5	0.8557
cosine_accuracy@10	0.9014
cosine_precision@1	0.6843
cosine_precision@3	0.2743
cosine_precision@5	0.1711
cosine_precision@10	0.0901
cosine_recall@1	0.6843
cosine_recall@3	0.8229
cosine_recall@5	0.8557
cosine_recall@10	0.9014
cosine_ndcg@10	0.794
cosine_mrr@10	0.7594
cosine_map@100	0.7636

Information Retrieval

• Dataset: dim_256
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.68
cosine_accuracy@3	0.8114
cosine_accuracy@5	0.85
cosine_accuracy@10	0.8943
cosine_precision@1	0.68
cosine_precision@3	0.2705
cosine_precision@5	0.17
cosine_precision@10	0.0894
cosine_recall@1	0.68
cosine_recall@3	0.8114
cosine_recall@5	0.85
cosine_recall@10	0.8943
cosine_ndcg@10	0.7889
cosine_mrr@10	0.755
cosine_map@100	0.7594

Information Retrieval

• Dataset: dim_128
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6571
cosine_accuracy@3	0.7943
cosine_accuracy@5	0.8343
cosine_accuracy@10	0.8886
cosine_precision@1	0.6571
cosine_precision@3	0.2648
cosine_precision@5	0.1669
cosine_precision@10	0.0889
cosine_recall@1	0.6571
cosine_recall@3	0.7943
cosine_recall@5	0.8343
cosine_recall@10	0.8886
cosine_ndcg@10	0.773
cosine_mrr@10	0.7361
cosine_map@100	0.7403

Information Retrieval

• Dataset: dim_64
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6186
cosine_accuracy@3	0.76
cosine_accuracy@5	0.8
cosine_accuracy@10	0.8657
cosine_precision@1	0.6186
cosine_precision@3	0.2533
cosine_precision@5	0.16
cosine_precision@10	0.0866
cosine_recall@1	0.6186
cosine_recall@3	0.76
cosine_recall@5	0.8
cosine_recall@10	0.8657
cosine_ndcg@10	0.7409
cosine_mrr@10	0.7013
cosine_map@100	0.7062

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: 8 tokens mean: 45.24 tokens max: 512 tokens	min: 9 tokens mean: 20.71 tokens max: 45 tokens

• Samples:

  positive	anchor
  Changes in Costs. Our costs are subject to fluctuations, particularly due to changes in commodity and input material prices, transportation costs, other broader inflationary impacts and our own productivity efforts. We have significant exposures to certain commodities and input materials, in particular certain oil-derived materials like resins and paper-based materials like pulp. Volatility in the market price of these commodities and input materials has a direct impact on our costs. Disruptions in our manufacturing, supply and distribution operations due to energy shortages, natural disasters, labor or freight constraints have impacted our costs and could do so in the future. New or increased legal or regulatory requirements, along with initiatives to meet our sustainability goals, could also result in increased costs due to higher material costs and investments in facilities and equipment. We strive to implement, achieve and sustain cost improvement plans, including supply chain optimization and general overhead and workforce optimization. Increased pricing in response to certain inflationary or cost increases may also offset portions of the cost impacts; however, such price increases may impact product consumption. If we are unable to manage cost impacts through pricing actions and consistent productivity improvements, it may adversely impact our net sales, gross margin, operating margin, net earnings and cash flows.	How did Procter & Gamble manage the fluctuations in costs, particularly related to commodities and input materials?
  As of October 1, 2023 we had ¥5 billion, or $33.5 million, of borrowings outstanding under these credit facilities.	How much was borrowed under the Japanese yen-denominated credit facilities as of October 1, 2023?
  AutoZone sells automotive hard parts, maintenance items, accessories and non-automotive products through www.autozone.com, and commercial customers can make purchases through www.autozonepro.com. Additionally, the ALLDATA brand of automotive diagnostic, repair, collision and shop management software is sold through www.alldata.com.	What online platforms does AutoZone use for selling automotive products and services?

• 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: False
• 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: False
• 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
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	dim_128_cosine_map@100	dim_256_cosine_map@100	dim_512_cosine_map@100	dim_64_cosine_map@100	dim_768_cosine_map@100
0.8122	10	1.5647	-	-	-	-	-
0.9746	12	-	0.7160	0.7404	0.7515	0.6797	0.7533
1.6244	20	0.6629	-	-	-	-	-
1.9492	24	-	0.7340	0.7582	0.7611	0.6996	0.7603
2.4365	30	0.4811	-	-	-	-	-
2.9239	36	-	0.7403	0.759	0.7638	0.7056	0.7646
3.2487	40	0.4046	-	-	-	-	-
3.8985	48	-	0.7403	0.7594	0.7636	0.7062	0.7647

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.10.12
• Sentence Transformers: 3.0.1
• Transformers: 4.41.2
• PyTorch: 2.3.0+cu121
• Accelerate: 0.31.0
• Datasets: 2.19.2
• 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}
}