willy-arison/bge-base-financial-willy2

BGE base Financial Matryoshka

This is a sentence-transformers model finetuned from BAAI/bge-base-en-v1.5 on the json 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:
  • json
• 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("willy-arison/bge-base-financial-willy2")
# Run inference
sentences = [
    'The topic of the first paper is "Fine-tuning gpt-3 for russian text summarization."',
    'What is the topic of the first paper mentioned in the text?',
    "What is the model's response when the date is changed to September 8, 2030?",
]
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	0.3333
cosine_accuracy@3	0.4667
cosine_accuracy@5	0.5333
cosine_accuracy@10	0.6
cosine_precision@1	0.3333
cosine_precision@3	0.1556
cosine_precision@5	0.1067
cosine_precision@10	0.06
cosine_recall@1	0.3333
cosine_recall@3	0.4667
cosine_recall@5	0.5333
cosine_recall@10	0.6
cosine_ndcg@10	0.4605
cosine_mrr@10	0.4167
cosine_map@100	0.4369

Information Retrieval

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

  {
      "truncate_dim": 512
  }
  

Metric	Value
cosine_accuracy@1	0.3667
cosine_accuracy@3	0.4667
cosine_accuracy@5	0.5
cosine_accuracy@10	0.6667
cosine_precision@1	0.3667
cosine_precision@3	0.1556
cosine_precision@5	0.1
cosine_precision@10	0.0667
cosine_recall@1	0.3667
cosine_recall@3	0.4667
cosine_recall@5	0.5
cosine_recall@10	0.6667
cosine_ndcg@10	0.4951
cosine_mrr@10	0.4429
cosine_map@100	0.4592

Information Retrieval

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

  {
      "truncate_dim": 256
  }
  

Metric	Value
cosine_accuracy@1	0.3
cosine_accuracy@3	0.4333
cosine_accuracy@5	0.5667
cosine_accuracy@10	0.6333
cosine_precision@1	0.3
cosine_precision@3	0.1444
cosine_precision@5	0.1133
cosine_precision@10	0.0633
cosine_recall@1	0.3
cosine_recall@3	0.4333
cosine_recall@5	0.5667
cosine_recall@10	0.6333
cosine_ndcg@10	0.4479
cosine_mrr@10	0.3897
cosine_map@100	0.4046

Information Retrieval

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

  {
      "truncate_dim": 128
  }
  

Metric	Value
cosine_accuracy@1	0.2667
cosine_accuracy@3	0.3667
cosine_accuracy@5	0.4667
cosine_accuracy@10	0.5667
cosine_precision@1	0.2667
cosine_precision@3	0.1222
cosine_precision@5	0.0933
cosine_precision@10	0.0567
cosine_recall@1	0.2667
cosine_recall@3	0.3667
cosine_recall@5	0.4667
cosine_recall@10	0.5667
cosine_ndcg@10	0.3964
cosine_mrr@10	0.3437
cosine_map@100	0.3564

Information Retrieval

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

  {
      "truncate_dim": 64
  }
  

Metric	Value
cosine_accuracy@1	0.2667
cosine_accuracy@3	0.3333
cosine_accuracy@5	0.4333
cosine_accuracy@10	0.5667
cosine_precision@1	0.2667
cosine_precision@3	0.1111
cosine_precision@5	0.0867
cosine_precision@10	0.0567
cosine_recall@1	0.2667
cosine_recall@3	0.3333
cosine_recall@5	0.4333
cosine_recall@10	0.5667
cosine_ndcg@10	0.3869
cosine_mrr@10	0.3324
cosine_map@100	0.3458

Training Details

Training Dataset

json

• Dataset: json
• Size: 264 training samples
• Columns: positive and anchor
• Approximate statistics based on the first 264 samples:

  	positive	anchor
  type	string	string
  details	min: 3 tokens mean: 15.37 tokens max: 61 tokens	min: 9 tokens mean: 18.22 tokens max: 37 tokens

• Samples:

  positive	anchor
  Hospital systems may wish to update an LLM with their current medical guidelines.	Give an example of a specific domain or industry that might want to update a language model with their own knowledge.
  The Gemini models struggle to learn a significant proportion of the data even after 20 or 30 epochs.	How do the Gemini models perform in learning the training data compared to the OpenAI models?
  Anthropic, Google, OpenAI	Which companies have contributed to the rapid iteration and evolution of Large Language Models?

• 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: 20
• lr_scheduler_type: cosine
• warmup_ratio: 0.1
• bf16: 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
• 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: 20
• 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: 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: 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: 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
1.0	1	-	0.3407	0.3622	0.3385	0.3376	0.3703
2.0	2	-	0.3739	0.3652	0.3634	0.3429	0.3613
3.0	3	-	0.3742	0.3666	0.3584	0.3495	0.3504
4.0	4	-	0.3838	0.3666	0.3580	0.3480	0.3488
1.0	1	-	0.3742	0.3666	0.3584	0.3495	0.3504
2.0	2	-	0.3939	0.3767	0.3611	0.3626	0.3302
3.0	3	-	0.3844	0.3896	0.3873	0.3589	0.3388
4.0	4	-	0.3885	0.4011	0.3989	0.3581	0.3515
5.0	5	-	0.3790	0.4024	0.3984	0.3658	0.3563
6.0	6	-	0.3815	0.4047	0.3998	0.3904	0.3573
7.0	7	-	0.3868	0.4164	0.4038	0.4024	0.3512
8.0	8	-	0.3971	0.4224	0.4047	0.3933	0.3518
9.0	9	-	0.3971	0.4224	0.4062	0.3933	0.3536
10.0	10	69.6696	0.3971	0.4353	0.4062	0.4056	0.3579
1.0	1	-	0.3971	0.4353	0.4062	0.4056	0.3579
2.0	2	-	0.3885	0.4454	0.4181	0.3947	0.3558
3.0	3	-	0.4141	0.4482	0.4424	0.4020	0.3584
4.0	4	-	0.4215	0.4502	0.4443	0.3854	0.3571
5.0	5	-	0.4402	0.4596	0.4366	0.3874	0.3671
6.0	6	-	0.4402	0.4500	0.4372	0.3901	0.3645
7.0	7	-	0.4551	0.4557	0.4390	0.4045	0.3489
8.0	8	-	0.4658	0.4637	0.4390	0.3865	0.3487
9.0	9	-	0.4658	0.4741	0.4586	0.3827	0.3587
10.0	10	41.2785	0.4590	0.4741	0.4510	0.3941	0.3600
11.0	11	-	0.4590	0.4741	0.4515	0.3841	0.3604
12.0	12	-	0.4605	0.4802	0.4556	0.3864	0.3945
13.0	13	-	0.4590	0.4792	0.4433	0.3864	0.3768
14.0	14	-	0.4595	0.4844	0.4447	0.3964	0.3812
15.0	15	-	0.4605	0.4945	0.4447	0.3974	0.3869
16.0	16	-	0.4611	0.4951	0.4465	0.3974	0.3861
17.0	17	-	0.4605	0.4951	0.4465	0.3974	0.3869
18.0	18	-	0.4605	0.4955	0.4588	0.4097	0.3905
19.0	19	-	0.4605	0.4951	0.4479	0.3974	0.3869
20.0	20	24.2435	0.4605	0.4951	0.4479	0.3964	0.3869

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.11.13
• Sentence Transformers: 4.1.0
• Transformers: 4.52.4
• PyTorch: 2.6.0+cu124
• Accelerate: 1.8.1
• Datasets: 3.6.0
• 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}
}