tomaarsen/distilroberta-base-nli-matryoshka-v3

SentenceTransformer based on distilbert/distilroberta-base

This is a sentence-transformers model finetuned from distilbert/distilroberta-base on the sentence-transformers/all-nli 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
• Base model: distilbert/distilroberta-base
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 tokens
• Similarity Function: Cosine Similarity
• Training Dataset:
  • sentence-transformers/all-nli
• 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': 512, '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-nli-matryoshka-v3")
# Run inference
sentences = [
    'A man shoots a man.',
    'A man is shooting off guns.',
    'A man is erasing a chalk board.',
]
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-768
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8481
spearman_cosine	0.8519
pearson_manhattan	0.8393
spearman_manhattan	0.8385
pearson_euclidean	0.841
spearman_euclidean	0.8402
pearson_dot	0.7784
spearman_dot	0.778
pearson_max	0.8481
spearman_max	0.8519

Semantic Similarity

• Dataset: sts-dev-512
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8481
spearman_cosine	0.8524
pearson_manhattan	0.8386
spearman_manhattan	0.8377
pearson_euclidean	0.8402
spearman_euclidean	0.8395
pearson_dot	0.7712
spearman_dot	0.7713
pearson_max	0.8481
spearman_max	0.8524

Semantic Similarity

• Dataset: sts-dev-256
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8421
spearman_cosine	0.8488
pearson_manhattan	0.8313
spearman_manhattan	0.8316
pearson_euclidean	0.8333
spearman_euclidean	0.8335
pearson_dot	0.7446
spearman_dot	0.745
pearson_max	0.8421
spearman_max	0.8488

Semantic Similarity

• Dataset: sts-dev-128
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8347
spearman_cosine	0.8445
pearson_manhattan	0.8241
spearman_manhattan	0.8248
pearson_euclidean	0.8254
spearman_euclidean	0.8262
pearson_dot	0.7084
spearman_dot	0.7093
pearson_max	0.8347
spearman_max	0.8445

Semantic Similarity

• Dataset: sts-dev-64
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8201
spearman_cosine	0.8352
pearson_manhattan	0.8032
spearman_manhattan	0.8047
pearson_euclidean	0.806
spearman_euclidean	0.8072
pearson_dot	0.636
spearman_dot	0.6389
pearson_max	0.8201
spearman_max	0.8352

Semantic Similarity

• Dataset: sts-test-768
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8262
spearman_cosine	0.8298
pearson_manhattan	0.8104
spearman_manhattan	0.8033
pearson_euclidean	0.8114
spearman_euclidean	0.8048
pearson_dot	0.7351
spearman_dot	0.7223
pearson_max	0.8262
spearman_max	0.8298

Semantic Similarity

• Dataset: sts-test-512
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8265
spearman_cosine	0.8303
pearson_manhattan	0.8092
spearman_manhattan	0.8022
pearson_euclidean	0.81
spearman_euclidean	0.8034
pearson_dot	0.7239
spearman_dot	0.7141
pearson_max	0.8265
spearman_max	0.8303

Semantic Similarity

• Dataset: sts-test-256
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8248
spearman_cosine	0.8305
pearson_manhattan	0.8012
spearman_manhattan	0.7951
pearson_euclidean	0.8028
spearman_euclidean	0.7974
pearson_dot	0.7011
spearman_dot	0.6946
pearson_max	0.8248
spearman_max	0.8305

Semantic Similarity

• Dataset: sts-test-128
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8206
spearman_cosine	0.8284
pearson_manhattan	0.7932
spearman_manhattan	0.7878
pearson_euclidean	0.7947
spearman_euclidean	0.7891
pearson_dot	0.6618
spearman_dot	0.6586
pearson_max	0.8206
spearman_max	0.8284

Semantic Similarity

• Dataset: sts-test-64
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.8119
spearman_cosine	0.8241
pearson_manhattan	0.7761
spearman_manhattan	0.7738
pearson_euclidean	0.7777
spearman_euclidean	0.7746
pearson_dot	0.5934
spearman_dot	0.5884
pearson_max	0.8119
spearman_max	0.8241

Training Details

Training Dataset

sentence-transformers/all-nli

• Dataset: sentence-transformers/all-nli at 65dd388
• 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: 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
  }
  

Evaluation Dataset

sentence-transformers/stsb

• Dataset: sentence-transformers/stsb at ab7a5ac
• Size: 1,500 evaluation samples
• Columns: sentence1, sentence2, and score
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	score
  type	string	string	float
  details	min: 5 tokens mean: 15.0 tokens max: 44 tokens	min: 6 tokens mean: 14.99 tokens max: 61 tokens	min: 0.0 mean: 0.47 max: 1.0

• Samples:

  sentence1	sentence2	score
  A man with a hard hat is dancing.	A man wearing a hard hat is dancing.	1.0
  A young child is riding a horse.	A child is riding a horse.	0.95
  A man is feeding a mouse to a snake.	The man is feeding a mouse to the snake.	1.0

• 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: 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

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: proportional

Training Logs

Epoch	Step	Training Loss	loss	sts-dev-128_spearman_cosine	sts-dev-256_spearman_cosine	sts-dev-512_spearman_cosine	sts-dev-64_spearman_cosine	sts-dev-768_spearman_cosine	sts-test-128_spearman_cosine	sts-test-256_spearman_cosine	sts-test-512_spearman_cosine	sts-test-64_spearman_cosine	sts-test-768_spearman_cosine
0.0229	100	19.9245	11.3900	0.7772	0.7998	0.8049	0.7902	0.7919	-	-	-	-	-
0.0459	200	10.6055	11.1510	0.7809	0.7996	0.8055	0.7954	0.7954	-	-	-	-	-
0.0688	300	9.6389	11.1229	0.7836	0.8029	0.8114	0.7923	0.8083	-	-	-	-	-
0.0918	400	8.6917	11.0299	0.7976	0.8117	0.8142	0.8002	0.8087	-	-	-	-	-
0.1147	500	8.3064	11.3586	0.7895	0.8058	0.8120	0.7978	0.8065	-	-	-	-	-
0.1376	600	7.8026	11.5047	0.7876	0.8015	0.8065	0.7934	0.8016	-	-	-	-	-
0.1606	700	7.9978	11.5823	0.7944	0.8067	0.8072	0.7994	0.8045	-	-	-	-	-
0.1835	800	6.9249	11.5862	0.7945	0.8054	0.8085	0.8012	0.8033	-	-	-	-	-
0.2065	900	7.1059	11.2365	0.7895	0.8035	0.8072	0.7956	0.8031	-	-	-	-	-
0.2294	1000	6.5483	11.3770	0.7853	0.7994	0.8039	0.7894	0.8024	-	-	-	-	-
0.2524	1100	6.6684	11.5038	0.7968	0.8087	0.8115	0.8002	0.8065	-	-	-	-	-
0.2753	1200	6.4661	11.4057	0.7980	0.8082	0.8103	0.8057	0.8070	-	-	-	-	-
0.2982	1300	6.501	11.2521	0.7974	0.8100	0.8111	0.8025	0.8079	-	-	-	-	-
0.3212	1400	6.0769	11.1458	0.7971	0.8103	0.8124	0.7982	0.8082	-	-	-	-	-
0.3441	1500	6.1919	11.3180	0.8039	0.8129	0.8144	0.8094	0.8098	-	-	-	-	-
0.3671	1600	5.8213	11.6196	0.7924	0.8072	0.8090	0.8003	0.8012	-	-	-	-	-
0.3900	1700	5.534	11.0700	0.7979	0.8104	0.8132	0.8028	0.8101	-	-	-	-	-
0.4129	1800	5.7536	11.0916	0.7934	0.8087	0.8149	0.8008	0.8085	-	-	-	-	-
0.4359	1900	5.3778	11.2658	0.7942	0.8084	0.8104	0.7980	0.8049	-	-	-	-	-
0.4588	2000	5.4925	11.4851	0.7932	0.8062	0.8086	0.7932	0.8057	-	-	-	-	-
0.4818	2100	5.3125	11.4833	0.7987	0.8119	0.8154	0.8012	0.8124	-	-	-	-	-
0.5047	2200	5.1914	11.2848	0.7784	0.7971	0.8037	0.7911	0.8004	-	-	-	-	-
0.5276	2300	5.2921	11.5364	0.7698	0.7910	0.7974	0.7839	0.7900	-	-	-	-	-
0.5506	2400	5.288	11.3944	0.7873	0.8011	0.8051	0.7877	0.8003	-	-	-	-	-
0.5735	2500	5.3697	11.4532	0.7949	0.8077	0.8111	0.7955	0.8069	-	-	-	-	-
0.5965	2600	5.1521	11.2788	0.7973	0.8095	0.8130	0.7940	0.8088	-	-	-	-	-
0.6194	2700	5.2316	11.2472	0.7948	0.8077	0.8102	0.7939	0.8053	-	-	-	-	-
0.6423	2800	5.2599	11.4171	0.7882	0.8029	0.8065	0.7888	0.8019	-	-	-	-	-
0.6653	2900	5.4052	11.4026	0.7871	0.8005	0.8021	0.7833	0.7985	-	-	-	-	-
0.6882	3000	5.3474	11.2084	0.7895	0.8047	0.8079	0.7928	0.8050	-	-	-	-	-
0.7112	3100	5.0336	11.3999	0.8023	0.8150	0.8182	0.8024	0.8168	-	-	-	-	-
0.7341	3200	5.2496	11.2307	0.8015	0.8137	0.8167	0.8000	0.8140	-	-	-	-	-
0.7571	3300	3.8712	10.9468	0.8396	0.8440	0.8471	0.8284	0.8479	-	-	-	-	-
0.7800	3400	2.7068	10.9292	0.8414	0.8453	0.8489	0.8305	0.8497	-	-	-	-	-
0.8029	3500	2.3418	10.8626	0.8427	0.8467	0.8504	0.8322	0.8504	-	-	-	-	-
0.8259	3600	2.2419	10.9065	0.8421	0.8467	0.8504	0.8320	0.8502	-	-	-	-	-
0.8488	3700	2.125	10.9517	0.8424	0.8472	0.8509	0.8324	0.8510	-	-	-	-	-
0.8718	3800	1.9942	11.0142	0.8438	0.8482	0.8519	0.8337	0.8517	-	-	-	-	-
0.8947	3900	2.031	10.9662	0.8433	0.8480	0.8519	0.8340	0.8515	-	-	-	-	-
0.9176	4000	1.9734	11.0054	0.8452	0.8495	0.8531	0.8354	0.8528	-	-	-	-	-
0.9406	4100	1.9468	11.0183	0.8447	0.8490	0.8526	0.8348	0.8522	-	-	-	-	-
0.9635	4200	1.9008	11.0154	0.8445	0.8485	0.8521	0.8352	0.8517	-	-	-	-	-
0.9865	4300	1.8511	10.9966	0.8445	0.8488	0.8524	0.8352	0.8519	-	-	-	-	-
1.0	4359	-	-	-	-	-	-	-	0.8284	0.8305	0.8303	0.8241	0.8298

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.346 kWh
• Carbon Emitted: 0.134 kg of CO2
• Hours Used: 1.296 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",
}

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}
}