BGE base PatentMatch Matryoshka

This is a sentence-transformers model finetuned from BAAI/bge-base-en-v1.5 on the bhlim/patentmatch_for_finetuning 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 Sources

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("bhlim/bge-base-patentmatch")
# Run inference
sentences = [
    'Referring to FIG.32 a a sink device 3200 is designed to display thumbnail images in the metadata of contents received from source devices connected via an integrated wire interface.As mentioned in the foregoing description if a remote controller 3250 capable of outputting a pointing signal is situated within a region of a specific thumbnail image 3260 side information e.g.Amanda 1st album singer.Song etc.is displayed together.',
    'The method of any one of claims 8 to 12 wherein the requesting for the broadcast channel information comprises transmitting to the server image data obtained by capturing the content being reproduced by the display apparatus or audio data obtained by recording the content for a certain time.',
    'The electrode assembly of any one of the preceding claims wherein the first electrode comprises a substrate 113 wherein the first active material layer comprises active material layers 112 on both surfaces of the substrate and the ceramic layer comprises ceramic material layers 50 on both surfaces of the substrate.',
]
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

Metric Value
cosine_accuracy@1 0.0426
cosine_accuracy@3 0.1014
cosine_accuracy@5 0.1448
cosine_accuracy@10 0.232
cosine_precision@1 0.0426
cosine_precision@3 0.0338
cosine_precision@5 0.029
cosine_precision@10 0.0232
cosine_recall@1 0.0426
cosine_recall@3 0.1014
cosine_recall@5 0.1448
cosine_recall@10 0.232
cosine_ndcg@10 0.1217
cosine_mrr@10 0.0884
cosine_map@100 0.1014

Information Retrieval

Metric Value
cosine_accuracy@1 0.0422
cosine_accuracy@3 0.0935
cosine_accuracy@5 0.1429
cosine_accuracy@10 0.2245
cosine_precision@1 0.0422
cosine_precision@3 0.0312
cosine_precision@5 0.0286
cosine_precision@10 0.0225
cosine_recall@1 0.0422
cosine_recall@3 0.0935
cosine_recall@5 0.1429
cosine_recall@10 0.2245
cosine_ndcg@10 0.1182
cosine_mrr@10 0.0861
cosine_map@100 0.0996

Information Retrieval

Metric Value
cosine_accuracy@1 0.0403
cosine_accuracy@3 0.0916
cosine_accuracy@5 0.1397
cosine_accuracy@10 0.2198
cosine_precision@1 0.0403
cosine_precision@3 0.0305
cosine_precision@5 0.0279
cosine_precision@10 0.022
cosine_recall@1 0.0403
cosine_recall@3 0.0916
cosine_recall@5 0.1397
cosine_recall@10 0.2198
cosine_ndcg@10 0.1151
cosine_mrr@10 0.0835
cosine_map@100 0.0963

Information Retrieval

Metric Value
cosine_accuracy@1 0.0379
cosine_accuracy@3 0.086
cosine_accuracy@5 0.1318
cosine_accuracy@10 0.208
cosine_precision@1 0.0379
cosine_precision@3 0.0287
cosine_precision@5 0.0264
cosine_precision@10 0.0208
cosine_recall@1 0.0379
cosine_recall@3 0.086
cosine_recall@5 0.1318
cosine_recall@10 0.208
cosine_ndcg@10 0.1089
cosine_mrr@10 0.0791
cosine_map@100 0.0909

Information Retrieval

Metric Value
cosine_accuracy@1 0.0328
cosine_accuracy@3 0.0742
cosine_accuracy@5 0.1144
cosine_accuracy@10 0.1847
cosine_precision@1 0.0328
cosine_precision@3 0.0247
cosine_precision@5 0.0229
cosine_precision@10 0.0185
cosine_recall@1 0.0328
cosine_recall@3 0.0742
cosine_recall@5 0.1144
cosine_recall@10 0.1847
cosine_ndcg@10 0.096
cosine_mrr@10 0.0692
cosine_map@100 0.0802

Training Details

Training Dataset

bhlim/patentmatch_for_finetuning

  • Dataset: bhlim/patentmatch_for_finetuning at 8d60f21
  • Size: 10,136 training samples
  • Columns: positive and anchor
  • Approximate statistics based on the first 1000 samples:
    positive anchor
    type string string
    details
    • min: 5 tokens
    • mean: 136.61 tokens
    • max: 512 tokens
    • min: 12 tokens
    • mean: 76.35 tokens
    • max: 512 tokens
  • Samples:
    positive anchor
    Furthermore according to this liquid consuming apparatus if the decompression level acting on the liquid sensing chamber 21 of the liquid container 1 i.e.the pressure loss arising in the connecting passage between the liquid storage portion 7 and the liquid sensing chamber 21 due to the flow rate outflowing from the liquid storage portion 7 because of distension of the diaphragm pump through application of the external force when external force is applied in the direction of expansion of volume of the diaphragm pump 42 asdepicted in FIG.6 has been set to a low level if sufficient liquid is present in the liquid container 1 the liquid sensing chamber 21 will experience substantially no change in volume. The liquid cartridge according to any of claims 4 to 5 further comprising a ground terminal 175c 176c 177c positioned in the second line.
    It is highly desirable for tires to have good wet skid resistance low rolling resistance and good wear characteristics.It has traditionally been very difficult to improve a tires wear characteristics without sacrificing its wet skid resistance and traction characteristics.These properties depend to a great extent on the dynamic viscoelastic properties of the rubbers utilized in making the tire. The pneumatic tire of at least one of the previous claims wherein the rubber composition comprises from 5 to 20 phr of the oil and from 45 to 70 phr of the terpene phenol resin.
    Before setting the environment of the mobile communication terminal a user stores a multimedia message composed of different kinds of contents i.e.images sounds and texts.For example reference block 201 indicates a multimedia message composed of several images sounds and texts.The user can select an image A a sound A and a text A for environment setting elements of the mobile communication terminal from the contents of the multimedia message and construct a theme like in block 203 using the selected image A sound A and text A.The MPU 101 maps the contents of the theme to environment setting elements of the mobile communication terminal i.e.a background screen a ringtone and a user name like in block 205.The MPU 101 then sets the environment of the mobile communication terminal using the mapped elements like in block 207 thereby automatically and collectively changing the environment of the mobile communication terminal.Mapping information about mapping between the selected contents of the multimediamessage and the environment setting elements of the mobile communication terminal is stored in the flash RAM 107. A terminal for processing data comprising an output unit configured to output a chatting service window a receiving unit configured to receive a request for executing a chatting service and a first download request for downloading first data through the chatting service from a user and a controller configured to control to output the first data downloaded in response to the received first download request to a background screen of the chatting service window.
  • 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 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
0.5047 10 10.0459 - - - - -
0.9590 19 - 0.0849 0.0915 0.0939 0.0778 0.0966
1.0095 20 7.1373 - - - - -
1.5142 30 5.9969 - - - - -
1.9685 39 - 0.0890 0.0965 0.1007 0.0795 0.1012
2.0189 40 5.2984 - - - - -
2.5237 50 4.884 - - - - -
2.9779 59 - 0.091 0.0967 0.099 0.0801 0.1013
3.0284 60 4.6633 - - - - -
3.5331 70 4.5226 - - - - -
3.8360 76 - 0.0909 0.0963 0.0996 0.0802 0.1014
  • The bold row denotes the saved checkpoint.

Framework Versions

  • Python: 3.10.12
  • Sentence Transformers: 3.0.1
  • Transformers: 4.41.2
  • PyTorch: 2.1.2+cu121
  • Accelerate: 0.32.1
  • 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}
}
Downloads last month
25
Safetensors
Model size
109M params
Tensor type
F32
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for bhlim/bge-base-patentmatch

Finetuned
(307)
this model

Dataset used to train bhlim/bge-base-patentmatch

Evaluation results