tomaarsen/distilroberta-base-paraphrases-multi

SentenceTransformer based on distilbert/distilroberta-base

This is a sentence-transformers model finetuned from distilbert/distilroberta-base on the all-nli, sentence-compression, simple-wiki, altlex, quora-duplicates, coco-captions, flickr30k-captions, yahoo-answers and stack-exchange 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: distilbert/distilroberta-base
• Maximum Sequence Length: 128 tokens
• Output Dimensionality: 768 tokens
• Similarity Function: Cosine Similarity
• Training Datasets:
  • all-nli
  • sentence-compression
  • simple-wiki
  • altlex
  • quora-duplicates
  • coco-captions
  • flickr30k-captions
  • yahoo-answers
  • stack-exchange
• 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: RobertaModel 
  (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/distilroberta-base-paraphrases-multi")
# Run inference
sentences = [
    'guy on a bike',
    'Man riding a bike',
    'A man cooks on a grill.',
]
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

Semantic Similarity

• Dataset: sts-dev
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8415
spearman_cosine	0.8452
pearson_manhattan	0.8502
spearman_manhattan	0.8517
pearson_euclidean	0.8535
spearman_euclidean	0.8555
pearson_dot	0.6505
spearman_dot	0.649
pearson_max	0.8535
spearman_max	0.8555

Semantic Similarity

• Dataset: sts-test
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8106
spearman_cosine	0.8145
pearson_manhattan	0.8225
spearman_manhattan	0.8131
pearson_euclidean	0.8255
spearman_euclidean	0.8165
pearson_dot	0.5911
spearman_dot	0.5761
pearson_max	0.8255
spearman_max	0.8165

Training Details

Training Datasets

all-nli

• Dataset: all-nli at cc6c526
• Size: 557,850 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: 10.38 tokens max: 45 tokens	min: 6 tokens mean: 12.8 tokens max: 39 tokens	min: 6 tokens mean: 13.4 tokens max: 50 tokens

• Samples:

  anchor	positive	negative
  A person on a horse jumps over a broken down airplane.	A person is outdoors, on a horse.	A person is at a diner, ordering an omelette.
  Children smiling and waving at camera	There are children present	The kids are frowning
  A boy is jumping on skateboard in the middle of a red bridge.	The boy does a skateboarding trick.	The boy skates down the sidewalk.

• Loss: MultipleNegativesRankingLoss with these parameters:

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

sentence-compression

• Dataset: sentence-compression at 605bc91
• Size: 180,000 training samples
• Columns: text and simplified
• Approximate statistics based on the first 1000 samples:

  	text	simplified
  type	string	string
  details	min: 10 tokens mean: 33.13 tokens max: 126 tokens	min: 5 tokens mean: 11.13 tokens max: 29 tokens

• Samples:

  text	simplified
  The USHL completed an expansion draft on Monday as 10 players who were on the rosters of USHL teams during the 2009-10 season were selected by the League's two newest entries, the Muskegon Lumberjacks and Dubuque Fighting Saints.	USHL completes expansion draft
  Major League Baseball Commissioner Bud Selig will be speaking at St. Norbert College next month.	Bud Selig to speak at St. Norbert College
  It's fresh cherry time in Michigan and the best time to enjoy this delicious and nutritious fruit.	It's cherry time

• Loss: MultipleNegativesRankingLoss with these parameters:

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

simple-wiki

• Dataset: simple-wiki at 60fd9b4
• Size: 102,225 training samples
• Columns: text and simplified
• Approximate statistics based on the first 1000 samples:

  	text	simplified
  type	string	string
  details	min: 9 tokens mean: 35.19 tokens max: 128 tokens	min: 8 tokens mean: 29.1 tokens max: 128 tokens

• Samples:

  text	simplified
  The greatest example has been in his present job ( then , Minister for Foreign Affairs ) , where he has perforce concentrated on Anglo-Irish relations and , in particular the North ( i.e. , Northern Ireland ) .	The greatest example has been in his present job ( then , Minister for Foreign Affairs ) , where he has perforce concentrated on Anglo-Irish relations and , in particular Northern Ireland ( .
  His reputation rose further when opposition leaders under parliamentary privilege alleged that Taoiseach Charles Haughey , who in January 1982 had been Leader of the Opposition , had not merely rung the President 's Office but threatened to end the career of the army officer who took the call and who , on Hillery 's explicit instructions , had refused to put through the call to the President .	President Hillery refused to speak to any opposition party politicians , but when Charles Haughey , who was Leader of the Opposition , had rang the President 's Office he threatened to end the career of the army officer answered and refused on Hillery 's explicit orders to put the call through to the President .
  He considered returning to medicine , perhaps moving with his wife , Maeve ( also a doctor ) to Africa .	He thought about returning to medicine , perhaps moving with his wife , Maeve ( also a doctor ) to Africa .

• Loss: MultipleNegativesRankingLoss with these parameters:

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

altlex

• Dataset: altlex at 97eb209
• Size: 112,696 training samples
• Columns: text and simplified
• Approximate statistics based on the first 1000 samples:

  	text	simplified
  type	string	string
  details	min: 9 tokens mean: 31.8 tokens max: 121 tokens	min: 6 tokens mean: 26.49 tokens max: 114 tokens

• Samples:

  text	simplified
  A set of 31 guns , cast 1729-1749 by the first master founder at the Royal Foundry , later the Royal Arsenal , Woolwich , were used to fire salutes until 1907 , often for Queen Victoria , who was a frequent visitor .	A set of 31 guns , cast 1729-1749 by the first master founder at the Royal Foundry , later the Royal Arsenal , Woolwich , were used to fire salutes until 1907 , often for Queen Victoria who was a frequent visitor .
  In 1929 , the building became vacant , and was given to Prince Edward , Prince of Wales , by his father , King George V . This became the Prince 's chief residence and was used extensively by him for entertaining and as a country retreat .	In 1929 , the building became vacant , and was given to Prince Edward , the Prince of Wales by his father , King George V . This became the Prince 's chief residence , and was used extensively by the Prince for entertaining and as a country retreat .
  Additions included an octagon room in the north-east side , in which the King regularly had dinner .	Additions included an octagon room in the North-East side , where the King regularly had dinner .

• Loss: MultipleNegativesRankingLoss with these parameters:

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

quora-duplicates

• Dataset: quora-duplicates at 451a485
• Size: 101,762 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: 13.72 tokens max: 42 tokens	min: 6 tokens mean: 13.5 tokens max: 44 tokens	min: 4 tokens mean: 14.56 tokens max: 62 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"
  }
  

coco-captions

• Dataset: coco-captions at bd26018
• Size: 414,010 training samples
• Columns: caption1 and caption2
• Approximate statistics based on the first 1000 samples:

  	caption1	caption2
  type	string	string
  details	min: 10 tokens mean: 13.65 tokens max: 25 tokens	min: 10 tokens mean: 13.65 tokens max: 25 tokens

• Samples:

  caption1	caption2
  A clock that blends in with the wall hangs in a bathroom.	A very clean and well decorated empty bathroom
  A very clean and well decorated empty bathroom	A bathroom with a border of butterflies and blue paint on the walls above it.
  A bathroom with a border of butterflies and blue paint on the walls above it.	An angled view of a beautifully decorated bathroom.

• Loss: MultipleNegativesRankingLoss with these parameters:

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

flickr30k-captions

• Dataset: flickr30k-captions at 0ef0ce3
• Size: 158,881 training samples
• Columns: caption1 and caption2
• Approximate statistics based on the first 1000 samples:

  	caption1	caption2
  type	string	string
  details	min: 6 tokens mean: 16.22 tokens max: 60 tokens	min: 6 tokens mean: 16.22 tokens max: 60 tokens

• Samples:

  caption1	caption2
  Two men in green shirts are standing in a yard.	Two young, White males are outside near many bushes.
  Two young, White males are outside near many bushes.	Two young guys with shaggy hair look at their hands while hanging out in the yard.
  Two young guys with shaggy hair look at their hands while hanging out in the yard.	A man in a blue shirt standing in a garden.

• Loss: MultipleNegativesRankingLoss with these parameters:

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

yahoo-answers

• Dataset: yahoo-answers at 93b3605
• Size: 599,417 training samples
• Columns: question and answer
• Approximate statistics based on the first 1000 samples:

  	question	answer
  type	string	string
  details	min: 12 tokens mean: 52.48 tokens max: 128 tokens	min: 13 tokens mean: 83.5 tokens max: 128 tokens

• Samples:

  question	answer
  why doesn't an optical mouse work on a glass table? or even on some surfaces?	why doesn't an optical mouse work on a glass table? Optical mice use an LED and a camera to rapidly capture images of the surface beneath the mouse. The infomation from the camera is analyzed by a DSP (Digital Signal Processor) and used to detect imperfections in the underlying surface and determine motion. Some materials, such as glass, mirrors or other very shiny, uniform surfaces interfere with the ability of the DSP to accurately analyze the surface beneath the mouse. \nSince glass is transparent and very uniform, the mouse is unable to pick up enough imperfections in the underlying surface to determine motion. Mirrored surfaces are also a problem, since they constantly reflect back the same image, causing the DSP not to recognize motion properly. When the system is unable to see surface changes associated with movement, the mouse will not work properly.
  What is the best off-road motorcycle trail ? long-distance trail throughout CA	What is the best off-road motorcycle trail ? i hear that the mojave road is amazing! \nsearch for it online.
  What is Trans Fat? How to reduce that? I heard that tras fat is bad for the body. Why is that? Where can we find it in our daily food?	What is Trans Fat? How to reduce that? Trans fats occur in manufactured foods during the process of partial hydrogenation, when hydrogen gas is bubbled through vegetable oil to increase shelf life and stabilize the original polyunsatured oil. The resulting fat is similar to saturated fat, which raises "bad" LDL cholesterol and can lead to clogged arteries and heart disease. \nUntil very recently, food labels were not required to list trans fats, and this health risk remained hidden to consumers. In early July, FDA regulations changed, and food labels will soon begin identifying trans fat content in processed foods.

• Loss: MultipleNegativesRankingLoss with these parameters:

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

stack-exchange

• Dataset: stack-exchange at 1c9657a
• Size: 304,525 training samples
• Columns: title1 and title2
• Approximate statistics based on the first 1000 samples:

  	title1	title2
  type	string	string
  details	min: 5 tokens mean: 15.04 tokens max: 63 tokens	min: 5 tokens mean: 15.91 tokens max: 80 tokens

• Samples:

  title1	title2
  what is the advantage of using the GPU rendering options in Android?	Can anyone explain all these Developer Options?
  Blank video when converting uncompressed AVI files with ffmpeg	FFmpeg lossy compression problems
  URL Rewriting of a query string in php	How to create friendly URL in php?

• Loss: MultipleNegativesRankingLoss with these parameters:

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

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 128
• per_device_eval_batch_size: 128
• num_train_epochs: 1
• warmup_ratio: 0.1
• fp16: True
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: round_robin

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: False
• per_device_train_batch_size: 128
• per_device_eval_batch_size: 128
• 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: round_robin

Training Logs

Epoch	Step	Training Loss	sts-dev_spearman_cosine	sts-test_spearman_cosine
0.0140	100	3.739	-	-
0.0279	200	1.1317	-	-
0.0419	300	0.9645	-	-
0.0558	400	0.9053	-	-
0.0698	500	0.8889	-	-
0.0838	600	0.8741	-	-
0.0977	700	0.8329	-	-
0.1117	800	0.8331	-	-
0.1256	900	0.8241	-	-
0.1396	1000	0.7829	0.8460	-
0.1535	1100	0.7871	-	-
0.1675	1200	0.7521	-	-
0.1815	1300	0.7905	-	-
0.1954	1400	0.7531	-	-
0.2094	1500	0.7677	-	-
0.2233	1600	0.7745	-	-
0.2373	1700	0.7651	-	-
0.2513	1800	0.7712	-	-
0.2652	1900	0.7476	-	-
0.2792	2000	0.7814	0.8370	-
0.2931	2100	0.7536	-	-
0.3071	2200	0.7689	-	-
0.3210	2300	0.7656	-	-
0.3350	2400	0.7672	-	-
0.3490	2500	0.6921	-	-
0.3629	2600	0.6778	-	-
0.3769	2700	0.6844	-	-
0.3908	2800	0.6907	-	-
0.4048	2900	0.6881	-	-
0.4188	3000	0.6815	0.8372	-
0.4327	3100	0.6869	-	-
0.4467	3200	0.698	-	-
0.4606	3300	0.6868	-	-
0.4746	3400	0.7174	-	-
0.4886	3500	0.6714	-	-
0.5025	3600	0.6698	-	-
0.5165	3700	0.6838	-	-
0.5304	3800	0.6927	-	-
0.5444	3900	0.6628	-	-
0.5583	4000	0.6647	0.8367	-
0.5723	4100	0.6766	-	-
0.5863	4200	0.6987	-	-
0.6002	4300	0.6895	-	-
0.6142	4400	0.6571	-	-
0.6281	4500	0.66	-	-
0.6421	4600	0.6747	-	-
0.6561	4700	0.6495	-	-
0.6700	4800	0.6746	-	-
0.6840	4900	0.6575	-	-
0.6979	5000	0.6712	0.8454	-
0.7119	5100	0.6627	-	-
0.7259	5200	0.6538	-	-
0.7398	5300	0.6659	-	-
0.7538	5400	0.6551	-	-
0.7677	5500	0.6548	-	-
0.7817	5600	0.673	-	-
0.7956	5700	0.6805	-	-
0.8096	5800	0.6537	-	-
0.8236	5900	0.6826	-	-
0.8375	6000	0.7182	0.8370	-
0.8515	6100	0.7391	-	-
0.8654	6200	0.7006	-	-
0.8794	6300	0.6774	-	-
0.8934	6400	0.7076	-	-
0.9073	6500	0.6893	-	-
0.9213	6600	0.678	-	-
0.9352	6700	0.6703	-	-
0.9492	6800	0.675	-	-
0.9631	6900	0.6842	-	-
0.9771	7000	0.6909	0.8452	-
0.9911	7100	0.681	-	-
1.0	7164	-	-	0.8145

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.202 kWh
• Carbon Emitted: 0.079 kg of CO2
• Hours Used: 0.601 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}
}