metadata

tags:
  - sentence-transformers
  - sentence-similarity
  - feature-extraction
  - generated_from_trainer
  - dataset_size:4415131
  - loss:TripletLoss
base_model: FacebookAI/roberta-base
pipeline_tag: sentence-similarity
library_name: sentence-transformers
metrics:
  - cosine_accuracy
  - cosine_accuracy_threshold
  - cosine_f1
  - cosine_f1_threshold
  - cosine_precision
  - cosine_recall
  - cosine_ap
  - cosine_mcc
model-index:
  - name: SentenceTransformer based on FacebookAI/roberta-base
    results:
      - task:
          type: binary-classification
          name: Binary Classification
        dataset:
          name: litemb dev
          type: litemb-dev
        metrics:
          - type: cosine_accuracy
            value: 0.833
            name: Cosine Accuracy
          - type: cosine_accuracy_threshold
            value: 0.799780011177063
            name: Cosine Accuracy Threshold
          - type: cosine_f1
            value: 0.8324429334628461
            name: Cosine F1
          - type: cosine_f1_threshold
            value: 0.7916845083236694
            name: Cosine F1 Threshold
          - type: cosine_precision
            value: 0.8092540132200189
            name: Cosine Precision
          - type: cosine_recall
            value: 0.857
            name: Cosine Recall
          - type: cosine_ap
            value: 0.9126964494743037
            name: Cosine Ap
          - type: cosine_mcc
            value: 0.6561430124870038
            name: Cosine Mcc
      - task:
          type: binary-classification
          name: Binary Classification
        dataset:
          name: litemb test
          type: litemb-test
        metrics:
          - type: cosine_accuracy
            value: 0.8371
            name: Cosine Accuracy
          - type: cosine_accuracy_threshold
            value: 0.9183984994888306
            name: Cosine Accuracy Threshold
          - type: cosine_f1
            value: 0.8420254124786649
            name: Cosine F1
          - type: cosine_f1_threshold
            value: 0.9132623076438904
            name: Cosine F1 Threshold
          - type: cosine_precision
            value: 0.8005769924269744
            name: Cosine Precision
          - type: cosine_recall
            value: 0.888
            name: Cosine Recall
          - type: cosine_ap
            value: 0.9163489411155188
            name: Cosine Ap
          - type: cosine_mcc
            value: 0.6708115884030683
            name: Cosine Mcc

SentenceTransformer based on FacebookAI/roberta-base

This is a sentence-transformers model finetuned from FacebookAI/roberta-base on the csv 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: FacebookAI/roberta-base
Maximum Sequence Length: 512 tokens
Output Dimensionality: 768 dimensions
Similarity Function: Cosine Similarity
Training Dataset:
- csv

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("sentence_transformers_model_id")
# Run inference
sentences = [
    "The account of an expedition against Fort Christina deserves to be\nquoted in full, for it is an example of what war might be, full of\nexcitement, and exercise, and heroism, without danger to life. We take\nup the narrative at the moment when the Dutch host...',
    '"He stood by me all these years," he thought, "he taught me all I know,\nthough I fear I am still very young and an ignoramus. But he\'s tried\nhard I know to impart all his own special knowledge to me, and he\'s\ngiven me chances that many a young officer would give his ears for.\nRight!...',
]
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

Binary Classification

Datasets: litemb-dev and litemb-test
Evaluated with BinaryClassificationEvaluator

Metric	litemb-dev	litemb-test
cosine_accuracy	0.833	0.8371
cosine_accuracy_threshold	0.7998	0.9184
cosine_f1	0.8324	0.842
cosine_f1_threshold	0.7917	0.9133
cosine_precision	0.8093	0.8006
cosine_recall	0.857	0.888
cosine_ap	0.9127	0.9163
cosine_mcc	0.6561	0.6708

Training Details

Training Dataset

csv

Dataset: csv
Size: 4,415,131 training samples
Columns: anchor, positive, and negative

Approximate statistics based on the first 1000 samples:

	anchor	positive	negative
type	string	string	string
details	min: 447 tokens mean: 510.65 tokens max: 512 tokens	min: 450 tokens mean: 510.71 tokens max: 512 tokens	min: 455 tokens mean: 510.83 tokens max: 512 tokens

Samples:

anchor	positive	negative
"That was curious," remarked Trent. "I thought so, sir. But I recollected what I had heard about 'not a word to a soul,' and I concluded that this about a moonlight drive was intended to mislead." "What time was this?" "It would be about ten, sir, I should say. After speaking to me, Mr. Manderson waited until Mr. Marlowe had come down and brought round the car. He then went into the drawing-room, where Mrs. Manderson was." "Did that strike you as curious?" Martin looked down his nose. "If you ask me the question, sir," he said with reserve, "I had not known him enter that room since we came here this year. He preferred to sit in the library in the evenings. That evening he only remained with Mrs. Manderson for a few minutes. Then he and Mr. Marlowe started immediately." "You saw them start?" "Yes, sir. They took the direction of Bishopsbridge." "And you saw Mr. Manderson again later?" "After an hour or thereabouts, sir, in the library. That would have been about a quarter past eleven, ...	Sir James turned instantly to Mr. Figgis, whose pencil was poised over the paper. “Sigsbee Manderson has been murdered,” he began quickly and clearly, pacing the floor with his hands behind him. Mr. Figgis scratched down a line of shorthand with as much emotion as if he had been told that the day was fine—the pose of his craft. “He and his wife and two secretaries have been for the past fortnight at the house called White Gables, at Marlstone, near Bishopsbridge. He bought it four years ago. He and Mrs. Manderson have since spent a part of each summer there. Last night he went to bed about half-past eleven, just as usual. No one knows when he got up and left the house. He was not missed until this morning. About ten o’clock his body was found by a gardener. It was lying by a shed in the grounds. He was shot in the head, through the left eye. Death must have been instantaneous. The body was not robbed, but there were marks on the wrists which pointed to a struggle having taken place. Dr...	Holmes shook his head like a man who is far from being satisfied. “These are very deep waters,” said he; “pray go on with your narrative.” “Two years have passed since then, and my life has been until lately lonelier than ever. A month ago, however, a dear friend, whom I have known for many years, has done me the honor to ask my hand in marriage. His name is Armitage—Percy Armitage—the second son of Mr. Armitage, of Crane Water, near Reading. My step-father has offered no opposition to the match, and we are to be married in the course of the spring. Two days ago some repairs were started in the west wing of the building, and my bedroom wall has been pierced, so that I have had to move into the chamber in which my sister died, and to sleep in the very bed in which she slept. Imagine, then, my thrill of terror when last night, as I lay awake, thinking over her terrible fate, I suddenly heard in the silence of the night the low whistle which had been the herald of her own death. I sprang ...
'The condition of those blacks is assuredly better than that of the agricultural laborers in many parts of Europe. Their morality is far superior to that of the free negroes of the North; the planters encourage marriage, and thus endeavor to develop among them a sense of the family relation, with a view of attaching them to the domestic hearth, consequently to the family of the master. It will be then observed that in such a state of things the interests of the planter, in default of any other motive, promotes the advancement and well-being of the slave. Certainly, we believe it possible still to ameliorate their condition. It is with that view, even, that the South has labored for so long a time to prepare them for a higher civilization. 'In no part, perhaps, of the continent, regard being had to the population, do there exist men more eminent and gifted, with nobler or more generous sentiments, than in the Southern States. No co...	If we had clear and strong faith, our joy at the thought of a glorified spirit, however necessary its presence to us here, would transcend all our sorrows; the streaming beams of sunshine would irradiate our weeping; we should think more of his happiness than of our discomfort. Instead of departed spirits falling asleep, it is we who have a spirit of slumber. O that we might walk by faith with glorified spirits before the throne, instead of remanding them,--as it seems we sometimes would do, if we could,--to the ignorance and infirmity of our condition. Our feelings towards the departed are the same as towards other prohibited things. Many are continually seeking for pleasures which God has taken away, or is purposely withholding from them. Let any one look at the history of his feelings, and see if his state of mind be not one of perpetual expectation of some form of happiness yet to arrive; an ideal of bliss, some prefigured condition, in which contentment and peace are to abide; whi...	“And we? Now that we've fought and lied and sweated and stolen, and hated as only the disappointed strugglers in a bitter, dead little Western town know how to do, what have we got to show for it? Harvey Merrick wouldn't have given one sunset over your marshes for all you've got put together, and you know it. It's not for me to say why, in the inscrutable wisdom of God, a genius should ever have been called from this place of hatred and bitter waters; but I want this Boston man to know that the drivel he's been hearing here tonight is the only tribute any truly great man could ever have from such a lot of sick, side-tracked, burnt-dog, land-poor sharks as the here-present financiers of Sand City--upon which town may God have mercy!” The lawyer thrust out his hand to Steavens as he passed him, caught up his overcoat in the hall, and had left the house before the Grand Army man had had time to lift his ducked head and crane his long neck about at his fellows. Next day Jim Laird was drun...
When Cowper became an author he paid the highest respect to Mrs. Unwin as an instinctive critic, and called her his Lord Chamberlain, whose approbation was his sufficient licence for publication. Life in the Unwin family is thus described by the new inmate;--"As to amusements, I mean what the world calls such, we have none. The place indeed swarms with them; and cards and dancing are the professed business of almost all the gentle inhabitants of Huntingdon. We refuse to take part in them, or to be accessories to this way of murdering our time, and by so doing have acquired the name of Methodists. Having told you how we do not spend our time, I will next say how we do. We breakfast commonly between eight and nine; till eleven, we read either the scripture, or the sermons of some faithful preacher of those holy mysteries; at eleven we attend divine service, which is performed here twice every day, and from twelve to three we separate, and amuse ourselves as we please. During that in...	Peel’s Government having been overthrown on the question of the Corn Laws by a combination which the Duke of Wellington characterized with military frankness, of Tory Protectionists, Whigs, Radicals, and Irish Nationalists, the whole under Semitic influence, its chief, for the short remainder of his life, held himself aloof from the party fray, encouraging no new combination, and content with watching over the safety of his great fiscal reform; though, as Greville says, had the Premiership been put to the vote, Peel would have been elected by an overwhelming majority. His personal following, Peelites as they were called, Graham, Gladstone, Lincoln, Cardwell, Sidney Herbert, and the rest, remained suspended between the two great parties. When Disraeli had thrown over protection, as he meant from the beginning to do, the only barrier of principle between the Peelites and the Conservatives was removed. Overtures were made by the Conservative leader, Lord Derby, to Gladstone, whose immense...	"If you take my advice," said Stanley who was fighting his way towards some remote goal or other, "you'll take a little flyer on Dr. Rice. That's what I'm going to do. There's a fellow on the other side of the ring has him a point higher than anyone else." Dick, without having made up his mind as to his own betting or not betting, helped his companion in his struggle to get through the crowd. Desperate energy was necessary. There was never any time for apologies; elbows were pushed into sides, toes were trodden on, scarfs twisted and sleeve-links broken; no matter, there was money to be won and there was no time either to consider passing annoyances or the possibility of loss. "Ah," said Stanley, finally, as they found themselves in front of a black-board that had a figure "7" chalked to the left of the name Dr. Rice and a "3" to the right. "Here we are! Now then, what are you going to do?" He whipped out a twenty dollar bill and crumpled it carefully into the palm of his hand. Dick th...

Loss: TripletLoss with these parameters:

{
    "distance_metric": "TripletDistanceMetric.EUCLIDEAN",
    "triplet_margin": 0.5
}

Evaluation Dataset

csv

Dataset: csv
Size: 944,948 evaluation samples
Columns: anchor, positive, and negative

Approximate statistics based on the first 1000 samples:

	anchor	positive	negative
type	string	string	string
details	min: 420 tokens mean: 510.66 tokens max: 512 tokens	min: 432 tokens mean: 510.77 tokens max: 512 tokens	min: 424 tokens mean: 510.38 tokens max: 512 tokens

Loss: TripletLoss with these parameters:

{
    "distance_metric": "TripletDistanceMetric.EUCLIDEAN",
    "triplet_margin": 0.5
}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 35
per_device_eval_batch_size: 35
warmup_ratio: 0.1
fp16: True
load_best_model_at_end: True
batch_sampler: no_duplicates

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: steps
prediction_loss_only: True
per_device_train_batch_size: 35
per_device_eval_batch_size: 35
per_gpu_train_batch_size: None
per_gpu_eval_batch_size: None
gradient_accumulation_steps: 1
eval_accumulation_steps: None
torch_empty_cache_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: 3
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
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: 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: 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
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
include_for_metrics: []
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
eval_on_start: False
use_liger_kernel: False
eval_use_gather_object: False
average_tokens_across_devices: False
prompts: None
batch_sampler: no_duplicates
multi_dataset_batch_sampler: proportional

The bold row denotes the saved checkpoint.

Framework Versions

Python: 3.12.0
Sentence Transformers: 3.4.0.dev0
Transformers: 4.46.3
PyTorch: 2.5.1+cu121
Accelerate: 1.1.1
Datasets: 3.1.0
Tokenizers: 0.20.3

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}
}