Hugging Face's logo

lv12
/
esci-nomic-embed-text-v1_5_4

Sentence Similarity
sentence-transformers
Safetensors
nomic_bert
feature-extraction
Generated from Trainer
dataset_size:1600000
loss:TripletLoss
custom_code
Eval Results
Inference Endpoints
text-embeddings-inference
Model card Files Community
Edit model card

SentenceTransformer

This is a sentence-transformers model trained on the triplets 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
  • Maximum Sequence Length: 8192 tokens
  • Output Dimensionality: 768 tokens
  • Similarity Function: Cosine Similarity
  • Training Dataset:
    • triplets

Model Sources

Full Model Architecture 

SentenceTransformer(
  (0): Transformer({'max_seq_length': 8192, 'do_lower_case': False}) with Transformer model: NomicBertModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
)

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("lv12/esci-nomic-embed-text-v1_5_4")
# Run inference
sentences = [
    'search_query: karoke set 2 microphone for adults',
    'search_document: Starion KS829-B Bluetooth Karaoke Machine l Pedestal Design w/Light Show l Two Karaoke Microphones, Starion, Black',
    'search_document: EARISE T26 Portable Karaoke Machine Bluetooth Speaker with Wireless Microphone, Rechargeable PA System with FM Radio, Audio Recording, Remote Control, Supports TF Card/USB, Perfect for Party, EARISE, ',
]
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

Triplet

Metric Value
cosine_accuracy 0.7298
dot_accuracy 0.2832
manhattan_accuracy 0.7282
euclidean_accuracy 0.7299
max_accuracy 0.7299

Semantic Similarity

Metric Value
pearson_cosine 0.4148
spearman_cosine 0.3997
pearson_manhattan 0.3771
spearman_manhattan 0.3699
pearson_euclidean 0.3778
spearman_euclidean 0.3708
pearson_dot 0.3814
spearman_dot 0.3817
pearson_max 0.4148
spearman_max 0.3997

Information Retrieval

Metric Value
cosine_accuracy@10 0.967
cosine_precision@10 0.6951
cosine_recall@10 0.6217
cosine_ndcg@10 0.83
cosine_mrr@10 0.9111
cosine_map@10 0.7758
dot_accuracy@10 0.946
dot_precision@10 0.6369
dot_recall@10 0.5693
dot_ndcg@10 0.7669
dot_mrr@10 0.8754
dot_map@10 0.6962

Training Details

Training Dataset

triplets

  • Dataset: triplets
  • Size: 1,600,000 training samples
  • Columns: anchor, positive, and negative
  • Approximate statistics based on the first 1000 samples:
    anchor positive negative
    type string string string
    details
    • min: 7 tokens
    • mean: 11.03 tokens
    • max: 39 tokens
    • min: 10 tokens
    • mean: 39.86 tokens
    • max: 104 tokens
    • min: 9 tokens
    • mean: 39.73 tokens
    • max: 159 tokens
  • Samples:
    anchor positive negative
    search_query: udt hydraulic fluid search_document: Triax Agra UTTO XL Synthetic Blend Tractor Transmission and Hydraulic Oil, 6,000 Hour Life, 50% Less wear, 36F Pour Point, Replaces All OEM Tractor Fluids (5 Gallon Pail), TRIAX, search_document: Shell Rotella T5 Synthetic Blend 15W-40 Diesel Engine Oil (1-Gallon, Case of 3), Shell Rotella,
    search_query: cheetah print iphone xs case search_document: iPhone Xs Case, iPhone Xs Case,Doowear Leopard Cheetah Protective Cover Shell For Girls Women,Slim Fit Anti Scratch Shockproof Soft TPU Bumper Flexible Rubber Gel Silicone Case for iPhone Xs / X-1, Ebetterr, 1 search_document: iPhone Xs & iPhone X Case, J.west Luxury Sparkle Bling Translucent Leopard Print Soft Silicone Phone Case Cover for Girls Women Flex Slim Design Pattern Drop Protective Case for iPhone Xs/x 5.8 inch, J.west, Leopard
    search_query: platform shoes search_document: Teva Women's Flatform Universal Platform Sandal, Black, 5 M US, Teva, Black search_document: Vans Women's Old Skool Platform Trainers, (Black/White Y28), 5 UK 38 EU, Vans, Black/White
  • Loss: TripletLoss with these parameters:
    {
    "distance_metric": "TripletDistanceMetric.COSINE",
    "triplet_margin": 0.8
}

Evaluation Dataset

triplets

  • Dataset: triplets
  • Size: 16,000 evaluation samples
  • Columns: anchor, positive, and negative
  • Approximate statistics based on the first 1000 samples:
    anchor positive negative
    type string string string
    details
    • min: 7 tokens
    • mean: 11.02 tokens
    • max: 29 tokens
    • min: 10 tokens
    • mean: 38.78 tokens
    • max: 87 tokens
    • min: 9 tokens
    • mean: 38.81 tokens
    • max: 91 tokens
  • Samples:
    anchor positive negative
    search_query: hogknobz search_document: Black 2014-2015 HDsmallPARTS/LocEzy Saddlebag Mounting Hardware Knobs are replacement/compatible for Saddlebag Quick Release Pins on Harley Davidson Touring Motorcycles Theft Deterrent, LocEzy, search_document: HANSWD Saddlebag Support Bars Brackets For SUZUKI YAMAHA KAWASAKI (Black), HANSWD, Black
    search_query: tile sticker key finder search_document: Tile Sticker (2020) 2-pack - Small, Adhesive Bluetooth Tracker, Item Locator and Finder for Remotes, Headphones, Gadgets and More, Tile, search_document: Tile Pro Combo (2017) - 2 Pack (1 x Sport, 1 x Style) - Discontinued by Manufacturer, Tile, Graphite/Gold
    search_query: adobe incense burner search_document: AM Incense Burner Frankincense Resin - Luxury Globe Charcoal Bakhoor Burners for Office & Home Decor (Brown), AM, Brown search_document: semli Large Incense Burner Backflow Incense Burner Holder Incense Stick Holder Home Office Decor, Semli,
  • Loss: TripletLoss with these parameters:
    {
    "distance_metric": "TripletDistanceMetric.COSINE",
    "triplet_margin": 0.8
}

Training Hyperparameters

Non-Default Hyperparameters

  • per_device_train_batch_size: 64
  • per_device_eval_batch_size: 16
  • gradient_accumulation_steps: 2
  • learning_rate: 1e-07
  • num_train_epochs: 5
  • lr_scheduler_type: polynomial
  • lr_scheduler_kwargs: {'lr_end': 1e-08, 'power': 2.0}
  • warmup_ratio: 0.05
  • dataloader_drop_last: True
  • dataloader_num_workers: 4
  • dataloader_prefetch_factor: 4
  • load_best_model_at_end: True
  • gradient_checkpointing: True
  • auto_find_batch_size: True
  • batch_sampler: no_duplicates

All Hyperparameters

Click to expand
  • overwrite_output_dir: False
  • do_predict: False
  • prediction_loss_only: True
  • per_device_train_batch_size: 64
  • per_device_eval_batch_size: 16
  • per_gpu_train_batch_size: None
  • per_gpu_eval_batch_size: None
  • gradient_accumulation_steps: 2
  • eval_accumulation_steps: None
  • learning_rate: 1e-07
  • weight_decay: 0.0
  • adam_beta1: 0.9
  • adam_beta2: 0.999
  • adam_epsilon: 1e-08
  • max_grad_norm: 1.0
  • num_train_epochs: 5
  • max_steps: -1
  • lr_scheduler_type: polynomial
  • lr_scheduler_kwargs: {'lr_end': 1e-08, 'power': 2.0}
  • warmup_ratio: 0.05
  • 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
  • no_cuda: False
  • use_cpu: False
  • use_mps_device: False
  • seed: 42
  • data_seed: None
  • jit_mode_eval: False
  • use_ipex: False
  • bf16: False
  • 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: True
  • dataloader_num_workers: 4
  • dataloader_prefetch_factor: 4
  • 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}
  • deepspeed: None
  • label_smoothing_factor: 0.0
  • optim: adamw_torch
  • 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: True
  • gradient_checkpointing_kwargs: None
  • include_inputs_for_metrics: False
  • fp16_backend: auto
  • push_to_hub_model_id: None
  • push_to_hub_organization: None
  • mp_parameters:
  • auto_find_batch_size: True
  • 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
  • batch_sampler: no_duplicates
  • multi_dataset_batch_sampler: proportional

Training Logs

Click to expand
Epoch Step Training Loss triplets loss cosine_accuracy cosine_map@10 spearman_cosine
0.0008 10 0.7505 - - - -
0.0016 20 0.7499 - - - -
0.0024 30 0.7524 - - - -
0.0032 40 0.7486 - - - -
0.004 50 0.7493 - - - -
0.0048 60 0.7476 - - - -
0.0056 70 0.7483 - - - -
0.0064 80 0.7487 - - - -
0.0072 90 0.7496 - - - -
0.008 100 0.7515 0.7559 0.7263 0.7684 0.3941
0.0088 110 0.7523 - - - -
0.0096 120 0.7517 - - - -
0.0104 130 0.7534 - - - -
0.0112 140 0.746 - - - -
0.012 150 0.7528 - - - -
0.0128 160 0.7511 - - - -
0.0136 170 0.7491 - - - -
0.0144 180 0.752 - - - -
0.0152 190 0.7512 - - - -
0.016 200 0.7513 0.7557 0.7259 0.7688 0.3942
0.0168 210 0.7505 - - - -
0.0176 220 0.7481 - - - -
0.0184 230 0.7516 - - - -
0.0192 240 0.7504 - - - -
0.02 250 0.7498 - - - -
0.0208 260 0.7506 - - - -
0.0216 270 0.7486 - - - -
0.0224 280 0.7471 - - - -
0.0232 290 0.7511 - - - -
0.024 300 0.7506 0.7553 0.7258 0.7692 0.3943
0.0248 310 0.7485 - - - -
0.0256 320 0.7504 - - - -
0.0264 330 0.7456 - - - -
0.0272 340 0.7461 - - - -
0.028 350 0.7496 - - - -
0.0288 360 0.7518 - - - -
0.0296 370 0.7514 - - - -
0.0304 380 0.7479 - - - -
0.0312 390 0.7507 - - - -
0.032 400 0.7511 0.7547 0.7258 0.7695 0.3945
0.0328 410 0.7491 - - - -
0.0336 420 0.7487 - - - -
0.0344 430 0.7496 - - - -
0.0352 440 0.7464 - - - -
0.036 450 0.7518 - - - -
0.0368 460 0.7481 - - - -
0.0376 470 0.7493 - - - -
0.0384 480 0.753 - - - -
0.0392 490 0.7475 - - - -
0.04 500 0.7498 0.7540 0.7262 0.7700 0.3948
0.0408 510 0.7464 - - - -
0.0416 520 0.7506 - - - -
0.0424 530 0.747 - - - -
0.0432 540 0.7462 - - - -
0.044 550 0.75 - - - -
0.0448 560 0.7522 - - - -
0.0456 570 0.7452 - - - -
0.0464 580 0.7475 - - - -
0.0472 590 0.7507 - - - -
0.048 600 0.7494 0.7531 0.7269 0.7707 0.3951
0.0488 610 0.7525 - - - -
0.0496 620 0.7446 - - - -
0.0504 630 0.7457 - - - -
0.0512 640 0.7462 - - - -
0.052 650 0.7478 - - - -
0.0528 660 0.7459 - - - -
0.0536 670 0.7465 - - - -
0.0544 680 0.7495 - - - -
0.0552 690 0.7513 - - - -
0.056 700 0.7445 0.7520 0.7274 0.7705 0.3954
0.0568 710 0.7446 - - - -
0.0576 720 0.746 - - - -
0.0584 730 0.7452 - - - -
0.0592 740 0.7459 - - - -
0.06 750 0.7419 - - - -
0.0608 760 0.7462 - - - -
0.0616 770 0.7414 - - - -
0.0624 780 0.7444 - - - -
0.0632 790 0.7419 - - - -
0.064 800 0.7438 0.7508 0.7273 0.7712 0.3957
0.0648 810 0.7503 - - - -
0.0656 820 0.7402 - - - -
0.0664 830 0.7435 - - - -
0.0672 840 0.741 - - - -
0.068 850 0.7386 - - - -
0.0688 860 0.7416 - - - -
0.0696 870 0.7473 - - - -
0.0704 880 0.7438 - - - -
0.0712 890 0.7458 - - - -
0.072 900 0.7446 0.7494 0.7279 0.7718 0.3961
0.0728 910 0.7483 - - - -
0.0736 920 0.7458 - - - -
0.0744 930 0.7473 - - - -
0.0752 940 0.7431 - - - -
0.076 950 0.7428 - - - -
0.0768 960 0.7385 - - - -
0.0776 970 0.7438 - - - -
0.0784 980 0.7406 - - - -
0.0792 990 0.7426 - - - -
0.08 1000 0.7372 0.7478 0.7282 0.7725 0.3965
0.0808 1010 0.7396 - - - -
0.0816 1020 0.7398 - - - -
0.0824 1030 0.7376 - - - -
0.0832 1040 0.7417 - - - -
0.084 1050 0.7408 - - - -
0.0848 1060 0.7415 - - - -
0.0856 1070 0.7468 - - - -
0.0864 1080 0.7427 - - - -
0.0872 1090 0.7371 - - - -
0.088 1100 0.7375 0.7460 0.7279 0.7742 0.3970
0.0888 1110 0.7434 - - - -
0.0896 1120 0.7441 - - - -
0.0904 1130 0.7378 - - - -
0.0912 1140 0.735 - - - -
0.092 1150 0.739 - - - -
0.0928 1160 0.7408 - - - -
0.0936 1170 0.7346 - - - -
0.0944 1180 0.7389 - - - -
0.0952 1190 0.7367 - - - -
0.096 1200 0.7358 0.7440 0.729 0.7747 0.3975
0.0968 1210 0.7381 - - - -
0.0976 1220 0.7405 - - - -
0.0984 1230 0.7348 - - - -
0.0992 1240 0.737 - - - -
0.1 1250 0.7393 - - - -
0.1008 1260 0.7411 - - - -
0.1016 1270 0.7359 - - - -
0.1024 1280 0.7276 - - - -
0.1032 1290 0.7364 - - - -
0.104 1300 0.7333 0.7418 0.7293 0.7747 0.3979
0.1048 1310 0.7367 - - - -
0.1056 1320 0.7352 - - - -
0.1064 1330 0.7333 - - - -
0.1072 1340 0.737 - - - -
0.108 1350 0.7361 - - - -
0.1088 1360 0.7299 - - - -
0.1096 1370 0.7339 - - - -
0.1104 1380 0.7349 - - - -
0.1112 1390 0.7318 - - - -
0.112 1400 0.7336 0.7394 0.7292 0.7749 0.3983
0.1128 1410 0.7326 - - - -
0.1136 1420 0.7317 - - - -
0.1144 1430 0.7315 - - - -
0.1152 1440 0.7321 - - - -
0.116 1450 0.7284 - - - -
0.1168 1460 0.7308 - - - -
0.1176 1470 0.7287 - - - -
0.1184 1480 0.727 - - - -
0.1192 1490 0.7298 - - - -
0.12 1500 0.7306 0.7368 0.7301 0.7755 0.3988
0.1208 1510 0.7269 - - - -
0.1216 1520 0.7299 - - - -
0.1224 1530 0.7256 - - - -
0.1232 1540 0.721 - - - -
0.124 1550 0.7274 - - - -
0.1248 1560 0.7251 - - - -
0.1256 1570 0.7248 - - - -
0.1264 1580 0.7244 - - - -
0.1272 1590 0.7275 - - - -
0.128 1600 0.7264 0.7339 0.7298 0.7756 0.3991
0.1288 1610 0.7252 - - - -
0.1296 1620 0.7287 - - - -
0.1304 1630 0.7263 - - - -
0.1312 1640 0.7216 - - - -
0.132 1650 0.7231 - - - -
0.1328 1660 0.728 - - - -
0.1336 1670 0.7309 - - - -
0.1344 1680 0.7243 - - - -
0.1352 1690 0.7239 - - - -
0.136 1700 0.7219 0.7309 0.7302 0.7768 0.3994
0.1368 1710 0.7212 - - - -
0.1376 1720 0.7217 - - - -
0.1384 1730 0.7118 - - - -
0.1392 1740 0.7226 - - - -
0.14 1750 0.7185 - - - -
0.1408 1760 0.7228 - - - -
0.1416 1770 0.7257 - - - -
0.1424 1780 0.7177 - - - -
0.1432 1790 0.722 - - - -
0.144 1800 0.712 0.7276 0.7307 0.7763 0.3997
0.1448 1810 0.7193 - - - -
0.1456 1820 0.7138 - - - -
0.1464 1830 0.7171 - - - -
0.1472 1840 0.7191 - - - -
0.148 1850 0.7172 - - - -
0.1488 1860 0.7168 - - - -
0.1496 1870 0.7111 - - - -
0.1504 1880 0.7203 - - - -
0.1512 1890 0.7095 - - - -
0.152 1900 0.7064 0.7240 0.7301 0.7762 0.3998
0.1528 1910 0.7147 - - - -
0.1536 1920 0.7098 - - - -
0.1544 1930 0.7193 - - - -
0.1552 1940 0.7096 - - - -
0.156 1950 0.7107 - - - -
0.1568 1960 0.7146 - - - -
0.1576 1970 0.7106 - - - -
0.1584 1980 0.7079 - - - -
0.1592 1990 0.7097 - - - -
0.16 2000 0.71 0.7202 0.7298 0.7758 0.3997

Framework Versions

  • Python: 3.10.12
  • Sentence Transformers: 3.0.1
  • Transformers: 4.38.2
  • PyTorch: 2.1.2+cu121
  • Accelerate: 0.27.2
  • Datasets: 2.19.1
  • Tokenizers: 0.15.2

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",
}

TripletLoss 

@misc{hermans2017defense,
    title={In Defense of the Triplet Loss for Person Re-Identification}, 
    author={Alexander Hermans and Lucas Beyer and Bastian Leibe},
    year={2017},
    eprint={1703.07737},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}
Downloads last month
16
Safetensors
Model size
137M params
Tensor type
F32
·
Inference API
Sentence Similarity
Examples
This model can be loaded on Inference API (serverless).

Evaluation results

View on Papers With Code