tessimago/bge-large-repmus-matryoshka

BGE base Financial Matryoshka

This is a sentence-transformers model finetuned from BAAI/bge-large-en-v1.5 on the json dataset. It maps sentences & paragraphs to a 1024-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-large-en-v1.5
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 1024 tokens
• 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': 1024, '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("tessimago/bge-large-repmus-matryoshka")
# Run inference
sentences = [
    'Sound funding decisions arise out of accurate assessments made of the SAR system. To measure the performance or effectiveness of a SAR system usually requires collecting information or statistics and establishing agreed-upon goals. All pertinent information should be collected, including where the system failed to perform as it should have; failures and successes provide valuable information in assessing effectiveness and determining means to improve. ',
    'What is required to measure the performance or effectiveness of a SAR system?',
    'What is the effect of decreasing track spacing on the area that can be searched?',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1024]

# 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	0.7632
cosine_accuracy@3	0.9123
cosine_accuracy@5	0.9386
cosine_accuracy@10	0.9912
cosine_precision@1	0.7632
cosine_precision@3	0.3041
cosine_precision@5	0.1877
cosine_precision@10	0.0991
cosine_recall@1	0.7632
cosine_recall@3	0.9123
cosine_recall@5	0.9386
cosine_recall@10	0.9912
cosine_ndcg@10	0.8801
cosine_mrr@10	0.8442
cosine_map@100	0.8449

Information Retrieval

• Dataset: dim_512
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.7456
cosine_accuracy@3	0.9211
cosine_accuracy@5	0.9386
cosine_accuracy@10	0.9912
cosine_precision@1	0.7456
cosine_precision@3	0.307
cosine_precision@5	0.1877
cosine_precision@10	0.0991
cosine_recall@1	0.7456
cosine_recall@3	0.9211
cosine_recall@5	0.9386
cosine_recall@10	0.9912
cosine_ndcg@10	0.8757
cosine_mrr@10	0.8383
cosine_map@100	0.8389

Information Retrieval

• Dataset: dim_256
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.7281
cosine_accuracy@3	0.8947
cosine_accuracy@5	0.9386
cosine_accuracy@10	0.9561
cosine_precision@1	0.7281
cosine_precision@3	0.2982
cosine_precision@5	0.1877
cosine_precision@10	0.0956
cosine_recall@1	0.7281
cosine_recall@3	0.8947
cosine_recall@5	0.9386
cosine_recall@10	0.9561
cosine_ndcg@10	0.8515
cosine_mrr@10	0.8167
cosine_map@100	0.8197

Information Retrieval

• Dataset: dim_128
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6842
cosine_accuracy@3	0.8596
cosine_accuracy@5	0.8947
cosine_accuracy@10	0.9386
cosine_precision@1	0.6842
cosine_precision@3	0.2865
cosine_precision@5	0.1789
cosine_precision@10	0.0939
cosine_recall@1	0.6842
cosine_recall@3	0.8596
cosine_recall@5	0.8947
cosine_recall@10	0.9386
cosine_ndcg@10	0.8139
cosine_mrr@10	0.7737
cosine_map@100	0.7778

Information Retrieval

• Dataset: dim_64
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.614
cosine_accuracy@3	0.7456
cosine_accuracy@5	0.8246
cosine_accuracy@10	0.8947
cosine_precision@1	0.614
cosine_precision@3	0.2485
cosine_precision@5	0.1649
cosine_precision@10	0.0895
cosine_recall@1	0.614
cosine_recall@3	0.7456
cosine_recall@5	0.8246
cosine_recall@10	0.8947
cosine_ndcg@10	0.748
cosine_mrr@10	0.7018
cosine_map@100	0.7074

Training Details

Training Dataset

json

• Dataset: json
• Size: 1,024 training samples
• Columns: positive and anchor
• Approximate statistics based on the first 1000 samples:

  	positive	anchor
  type	string	string
  details	min: 10 tokens mean: 133.58 tokens max: 512 tokens	min: 7 tokens mean: 17.7 tokens max: 39 tokens

• Samples:

  positive	anchor
  The debriefing helps to ensure that all survivors are rescued, to attend to the physical welfare of each survivor, and to obtain information which may assist and improve SAR services. Proper debriefing techniques include:– due care to avoid worsening a survivor’s condition by excessive debriefing;– careful assessment of the survivor’s statements if the survivor is frightened or excited;– use of a calm voice in questioning;– avoidance of suggesting the answers when obtaining facts; and– explaining that the information requested is important for the success of the SAR operation, and possibly for future SAR operations.	What are some proper debriefing techniques used in SAR services?
  Communicating with passengers is more difficult in remote areas where phone service may be inadequate or lacking. If phones do exist, calling the airline or shipping company may be the best way to check in and find out information. In more populated areas, local agencies may have an emergency evacuation plan or other useful plan that can be implemented.IE961E.indb 21 6/28/2013 10:29:55 AM	What is a good way to check in and find out information in remote areas where phone service may be inadequate or lacking?
  Voice communication is the basis of telemedical advice. It allows free dialogue and contributes to the human relationship, which is crucial to any medical consultation. Text messages are a useful complement to the voice telemedical advice and add the reliability of writing. Facsimile allows the exchange of pictures or diagrams, which help to identify a symptom, describe a lesion or the method of treatment. Digital data transmissions (photographs or electrocardiogram) provide an objective and potentially crucial addition to descriptive and subjective clinical data.	What are the types of communication methods used in telemedical advice?

• 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
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	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
1.0	2	0.7826	0.8163	0.8230	0.6761	0.8359
2.0	4	0.7739	0.8218	0.8282	0.6939	0.8459
3.0	6	0.7740	0.8223	0.8409	0.7072	0.8457
4.0	8	0.7778	0.8197	0.8389	0.7074	0.8449

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.10.14
• Sentence Transformers: 3.1.0
• Transformers: 4.41.2
• PyTorch: 2.1.2+cu121
• Accelerate: 0.34.2
• 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}
}