tomaarsen/stsb-distilbert-base-mnrl-cl-multi

SentenceTransformer based on sentence-transformers/stsb-distilbert-base

This is a sentence-transformers model finetuned from sentence-transformers/stsb-distilbert-base on the mnrl and cl datasets. 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: sentence-transformers/stsb-distilbert-base
• Maximum Sequence Length: 128 tokens
• Output Dimensionality: 768 tokens
• Similarity Function: Cosine Similarity
• Training Datasets:
  • mnrl
  • cl
• Language: en

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': 128, 'do_lower_case': False}) with Transformer model: DistilBertModel 
  (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("tomaarsen/stsb-distilbert-base-mnrl-cl-multi")
# Run inference
sentences = [
    'How fast is fast?',
    'How does light travel so fast?',
    'How do I copyright my books?',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings)
print(similarities.shape)
# [3, 3]

Evaluation

Metrics

Binary Classification

• Dataset: quora-duplicates
• Evaluated with BinaryClassificationEvaluator

Metric	Value
cosine_accuracy	0.846
cosine_accuracy_threshold	0.7969
cosine_f1	0.7791
cosine_f1_threshold	0.714
cosine_precision	0.6978
cosine_recall	0.882
cosine_ap	0.823
dot_accuracy	0.843
dot_accuracy_threshold	151.2908
dot_f1	0.7661
dot_f1_threshold	143.7784
dot_precision	0.7238
dot_recall	0.8137
dot_ap	0.7946
manhattan_accuracy	0.838
manhattan_accuracy_threshold	194.9912
manhattan_f1	0.7704
manhattan_f1_threshold	247.4978
manhattan_precision	0.6537
manhattan_recall	0.9379
manhattan_ap	0.815
euclidean_accuracy	0.841
euclidean_accuracy_threshold	9.0223
euclidean_f1	0.7704
euclidean_f1_threshold	11.3852
euclidean_precision	0.6463
euclidean_recall	0.9534
euclidean_ap	0.8153
max_accuracy	0.846
max_accuracy_threshold	194.9912
max_f1	0.7791
max_f1_threshold	247.4978
max_precision	0.7238
max_recall	0.9534
max_ap	0.823

Paraphrase Mining

• Dataset: quora-duplicates-dev
• Evaluated with ParaphraseMiningEvaluator

Metric	Value
average_precision	0.5889
f1	0.5762
precision	0.5478
recall	0.6077
threshold	0.7729

Information Retrieval

• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.963
cosine_accuracy@3	0.9906
cosine_accuracy@5	0.9944
cosine_accuracy@10	0.9982
cosine_precision@1	0.963
cosine_precision@3	0.4285
cosine_precision@5	0.2757
cosine_precision@10	0.1449
cosine_recall@1	0.83
cosine_recall@3	0.959
cosine_recall@5	0.9806
cosine_recall@10	0.9926
cosine_ndcg@10	0.9784
cosine_mrr@10	0.9772
cosine_map@100	0.9709
dot_accuracy@1	0.9514
dot_accuracy@3	0.9852
dot_accuracy@5	0.991
dot_accuracy@10	0.9968
dot_precision@1	0.9514
dot_precision@3	0.4247
dot_precision@5	0.2736
dot_precision@10	0.1446
dot_recall@1	0.8194
dot_recall@3	0.952
dot_recall@5	0.9756
dot_recall@10	0.9911
dot_ndcg@10	0.9715
dot_mrr@10	0.9693
dot_map@100	0.9617

Training Details

Training Datasets

mnrl

• Dataset: mnrl at 451a485
• Size: 100,000 training samples
• Columns: anchor, positive, and negative
• Approximate statistics based on the first 1000 samples:

  	anchor	positive	negative
  type	string	string	string
  details	min: 6 tokens mean: 13.85 tokens max: 42 tokens	min: 6 tokens mean: 13.65 tokens max: 44 tokens	min: 4 tokens mean: 14.76 tokens max: 64 tokens

• Samples:

  anchor	positive	negative
  Why in India do we not have one on one political debate as in USA?	Why cant we have a public debate between politicians in India like the one in US?	Can people on Quora stop India Pakistan debate? We are sick and tired seeing this everyday in bulk?
  What is OnePlus One?	How is oneplus one?	Why is OnePlus One so good?
  Does our mind control our emotions?	How do smart and successful people control their emotions?	How can I control my positive emotions for the people whom I love but they don't care about me?

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim"
  }
  

cl

• Dataset: cl at 451a485
• Size: 100,000 training samples
• Columns: sentence1, sentence2, and label
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	label
  type	string	string	int
  details	min: 6 tokens mean: 15.3 tokens max: 57 tokens	min: 6 tokens mean: 15.66 tokens max: 56 tokens	0: ~62.00% 1: ~38.00%

• Samples:

  sentence1	sentence2	label
  What is the step by step guide to invest in share market in india?	What is the step by step guide to invest in share market?	0
  What is the story of Kohinoor (Koh-i-Noor) Diamond?	What would happen if the Indian government stole the Kohinoor (Koh-i-Noor) diamond back?	0
  How can I increase the speed of my internet connection while using a VPN?	How can Internet speed be increased by hacking through DNS?	0

• Loss: ContrastiveLoss with these parameters:

  {
      "distance_metric": "SiameseDistanceMetric.COSINE_DISTANCE",
      "margin": 0.5,
      "size_average": true
  }
  

Evaluation Datasets

mnrl

• Dataset: mnrl at 451a485
• Size: 1,000 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: 13.84 tokens max: 43 tokens	min: 6 tokens mean: 13.8 tokens max: 38 tokens	min: 6 tokens mean: 14.71 tokens max: 56 tokens

• Samples:

  anchor	positive	negative
  Which programming language is best for developing low-end games?	What coding language should I learn first for making games?	I am entering the world of video game programming and want to know what language I should learn? Because there are so many languages ​​I do not know which one to start with. Can you recommend a language that's easy to learn and can be used with many platforms?
  Was it appropriate for Meryl Streep to use her Golden Globes speech to attack Donald Trump?	Should Meryl Streep be using her position to attack the president?	Why did Kelly Ann Conway say that Meryl Streep incited peoples worst feelings?
  Where can I found excellent commercial fridges in Sydney?	Where can I found impressive range of commercial fridges in Sydney?	What is the best grocery delivery service in Sydney?

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim"
  }
  

cl

• Dataset: cl at 451a485
• Size: 1,000 evaluation samples
• Columns: sentence1, sentence2, and label
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	label
  type	string	string	int
  details	min: 5 tokens mean: 15.59 tokens max: 59 tokens	min: 6 tokens mean: 15.65 tokens max: 76 tokens	0: ~63.40% 1: ~36.60%

• Samples:

  sentence1	sentence2	label
  What should I ask my friend to get from UK to India?	What is the process of getting a surgical residency in UK after completing MBBS from India?	0
  How can I learn hacking for free?	How can I learn to hack seriously?	1
  Which is the best website to learn programming language C++?	Which is the best website to learn C++ Programming language for free?	0

• Loss: ContrastiveLoss with these parameters:

  {
      "distance_metric": "SiameseDistanceMetric.COSINE_DISTANCE",
      "margin": 0.5,
      "size_average": true
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 64
• per_device_eval_batch_size: 64
• num_train_epochs: 1
• warmup_ratio: 0.1
• fp16: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: False
• per_device_train_batch_size: 64
• per_device_eval_batch_size: 64
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_steps: None
• learning_rate: 5e-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: 1
• max_steps: -1
• lr_scheduler_type: linear
• 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
• 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: True
• 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: False
• 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
• 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: None
• 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_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	cl loss	mnrl loss	cosine_map@100	quora-duplicates-dev_average_precision	quora-duplicates_max_ap
0	0	-	-	-	0.9245	0.4200	0.6890
0.0320	100	0.1634	-	-	-	-	-
0.0640	200	0.1206	-	-	-	-	-
0.0800	250	-	0.0190	0.1469	0.9530	0.5068	0.7354
0.0960	300	0.1036	-	-	-	-	-
0.1280	400	0.0836	-	-	-	-	-
0.1599	500	0.0918	0.0180	0.1008	0.9553	0.5259	0.7643
0.1919	600	0.0784	-	-	-	-	-
0.2239	700	0.0656	-	-	-	-	-
0.2399	750	-	0.0177	0.0905	0.9593	0.5305	0.7686
0.2559	800	0.0593	-	-	-	-	-
0.2879	900	0.0534	-	-	-	-	-
0.3199	1000	0.0612	0.0161	0.0736	0.9642	0.5512	0.7881
0.3519	1100	0.0572	-	-	-	-	-
0.3839	1200	0.06	-	-	-	-	-
0.3999	1250	-	0.0158	0.0641	0.9649	0.5567	0.7983
0.4159	1300	0.0565	-	-	-	-	-
0.4479	1400	0.0565	-	-	-	-	-
0.4798	1500	0.0475	0.0154	0.0578	0.9645	0.5614	0.8062
0.5118	1600	0.0596	-	-	-	-	-
0.5438	1700	0.0509	-	-	-	-	-
0.5598	1750	-	0.0150	0.0525	0.9674	0.5762	0.8092
0.5758	1800	0.0403	-	-	-	-	-
0.6078	1900	0.0431	-	-	-	-	-
0.6398	2000	0.0481	0.0150	0.0531	0.9689	0.5824	0.8128
0.6718	2100	0.05	-	-	-	-	-
0.7038	2200	0.0468	-	-	-	-	-
0.7198	2250	-	0.0146	0.0486	0.9684	0.5756	0.8195
0.7358	2300	0.0436	-	-	-	-	-
0.7678	2400	0.0409	-	-	-	-	-
0.7997	2500	0.0391	0.0145	0.0454	0.9705	0.5822	0.8190
0.8317	2600	0.0412	-	-	-	-	-
0.8637	2700	0.0373	-	-	-	-	-
0.8797	2750	-	0.0143	0.0451	0.9705	0.5889	0.8229
0.8957	2800	0.0428	-	-	-	-	-
0.9277	2900	0.0419	-	-	-	-	-
0.9597	3000	0.0376	0.0143	0.0435	0.9709	0.5889	0.8230
0.9917	3100	0.0366	-	-	-	-	-

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.084 kWh
• Carbon Emitted: 0.033 kg of CO2
• Hours Used: 0.399 hours

Training Hardware

• On Cloud: No
• GPU Model: 1 x NVIDIA GeForce RTX 3090
• CPU Model: 13th Gen Intel(R) Core(TM) i7-13700K
• RAM Size: 31.78 GB

Framework Versions

• Python: 3.11.6
• Sentence Transformers: 3.0.0.dev0
• Transformers: 4.41.0.dev0
• PyTorch: 2.3.0+cu121
• Accelerate: 0.26.1
• Datasets: 2.18.0
• 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",
}

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}
}

ContrastiveLoss

@inproceedings{hadsell2006dimensionality,
    author={Hadsell, R. and Chopra, S. and LeCun, Y.},
    booktitle={2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06)}, 
    title={Dimensionality Reduction by Learning an Invariant Mapping}, 
    year={2006},
    volume={2},
    number={},
    pages={1735-1742},
    doi={10.1109/CVPR.2006.100}
}