metadata

language:
  - multilingual
  - zh
  - ja
  - ar
  - ko
  - de
  - fr
  - es
  - pt
  - hi
  - id
  - it
  - tr
  - ru
  - bn
  - ur
  - mr
  - ta
  - vi
  - fa
  - pl
  - uk
  - nl
  - sv
  - he
  - sw
  - ps
library_name: sentence-transformers
tags:
  - sentence-transformers
  - sentence-similarity
  - feature-extraction
  - dataset_size:10K<n<100K
  - loss:CoSENTLoss
base_model: sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
metrics:
  - pearson_cosine
  - spearman_cosine
  - pearson_manhattan
  - spearman_manhattan
  - pearson_euclidean
  - spearman_euclidean
  - pearson_dot
  - spearman_dot
  - pearson_max
  - spearman_max
widget:
  - source_sentence: Is that wrong?
    sentences:
      - Is that such a terrible thing?
      - Kennedy korkunç bir savcıydı.
      - Tom bir davada tanıklık ediyordu.
  - source_sentence: Orada mıydılar?
    sentences:
      - Were they in there?
      - İlki ikincisini anlamlı kılar.
      - Alerji tedavisi gelişiyor.
  - source_sentence: He is not alone
    sentences:
      - It is not confusing
      - The Hawks were humanitarians.
      - Tom bir davada tanıklık ediyordu.
  - source_sentence: Yaptığın şey bu.
    sentences:
      - Onurlu işler yapıyorsunuz.
      - Weisberg azınlık adına konuştu.
      - Robert Ferrigno Kaliforniya'da doğdu.
  - source_sentence: Ben vatansızım.
    sentences:
      - I am stateless.
      - Kendi tekniğini tercih ediyor.
      - Mermiler camdan fırladı.
pipeline_tag: sentence-similarity
model-index:
  - name: >-
      SentenceTransformer based on
      sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2
    results:
      - task:
          type: semantic-similarity
          name: Semantic Similarity
        dataset:
          name: tr ling
          type: tr_ling
        metrics:
          - type: pearson_cosine
            value: 0.037604255015168134
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.04804112988506346
            name: Spearman Cosine
          - type: pearson_manhattan
            value: 0.034740275152181296
            name: Pearson Manhattan
          - type: spearman_manhattan
            value: 0.03769766156967754
            name: Spearman Manhattan
          - type: pearson_euclidean
            value: 0.03698411306484619
            name: Pearson Euclidean
          - type: spearman_euclidean
            value: 0.03903062430281842
            name: Spearman Euclidean
          - type: pearson_dot
            value: 0.0673696846368413
            name: Pearson Dot
          - type: spearman_dot
            value: 0.06818119362900125
            name: Spearman Dot
          - type: pearson_max
            value: 0.0673696846368413
            name: Pearson Max
          - type: spearman_max
            value: 0.06818119362900125
            name: Spearman Max

SentenceTransformer based on sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2

This is a sentence-transformers model finetuned from sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 on the MoritzLaurer/multilingual-nli-26lang-2mil7 dataset. It maps sentences & paragraphs to a 384-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/paraphrase-multilingual-MiniLM-L12-v2
Maximum Sequence Length: 128 tokens
Output Dimensionality: 384 tokens
Similarity Function: Cosine Similarity
Training Dataset:
- MoritzLaurer/multilingual-nli-26lang-2mil7
Languages: multilingual, zh, ja, ar, ko, de, fr, es, pt, hi, id, it, tr, ru, bn, ur, mr, ta, vi, fa, pl, uk, nl, sv, he, sw, ps

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: BertModel 
  (1): Pooling({'word_embedding_dimension': 384, '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 = [
    'Ben vatansızım.',
    'I am stateless.',
    'Kendi tekniğini tercih ediyor.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 384]

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

Evaluation

Metrics

Semantic Similarity

Dataset: tr_ling
Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.0376
spearman_cosine	0.048
pearson_manhattan	0.0347
spearman_manhattan	0.0377
pearson_euclidean	0.037
spearman_euclidean	0.039
pearson_dot	0.0674
spearman_dot	0.0682
pearson_max	0.0674
spearman_max	0.0682

Training Details

Training Dataset

MoritzLaurer/multilingual-nli-26lang-2mil7

Dataset: MoritzLaurer/multilingual-nli-26lang-2mil7 at 510a233
Size: 25,000 training samples
Columns: premise_original, hypothesis_original, score, sentence1, and sentence2

Approximate statistics based on the first 1000 samples:

	premise_original	hypothesis_original	score	sentence1	sentence2
type	string	string	int	string	string
details	min: 4 tokens mean: 29.3 tokens max: 107 tokens	min: 4 tokens mean: 15.62 tokens max: 40 tokens	0: ~34.50% 1: ~33.30% 2: ~32.20%	min: 4 tokens mean: 28.28 tokens max: 101 tokens	min: 4 tokens mean: 15.39 tokens max: 38 tokens

Samples:

premise_original	hypothesis_original	score	sentence1	sentence2
`N, the total number of LC50 values used in calculating the CV(%) varied with organism and toxicant because some data were rejected due to water hardness, lack of concentration measurements, and/or because some of the LC50s were not calculable.`	`Most discarded data was rejected due to water hardness.`	`1`	`N, CV'nin hesaplanmasında kullanılan LC50 değerlerinin toplam sayısı (%) organizma ve toksik madde ile çeşitlidir, çünkü bazı veriler su sertliği, konsantrasyon ölçümlerinin eksikliği ve / veya LC50'lerin bazıları hesaplanamaz olduğu için reddedilmiştir.`	`Atılan verilerin çoğu su sertliği nedeniyle reddedildi.`
`As the home of the Venus de Milo and Mona Lisa, the Louvre drew almost unmanageable crowds until President Mitterrand ordered its re-organization in the 1980s.`	`The Louvre is home of the Venus de Milo and Mona Lisa.`	`0`	`Venus de Milo ve Mona Lisa'nın evi olarak Louvre, Başkan Mitterrand'ın 1980'lerde yeniden düzenlenmesini emredene kadar neredeyse yönetilemez kalabalıklar çekti.`	`Louvre, Venus de Milo ve Mona Lisa'nın evidir.`
`A year ago, the wife of the Oxford don noticed that the pattern on Kleenex quilted tissue uncannily resembled the Penrose Arrowed Rhombi tilings pattern, which Sir Roger had invented--and copyrighted--in 1974.`	`It has been recently found out a similarity between the pattern on the recent Kleenex quilted tissue and the one of the Penrose Arrowed Rhombi tilings.`	`0`	`Bir yıl önce Oxford'un karısı, Kleenex kapitone dokudaki desenin 1974'te Sir Roger'ın icat ettiği -ve telif hakkı olan - Penrose Arrowed Rhombi tilings desenine benzediğini fark etti.`	`Yakın zamanda, son Kleenex kapitone dokudaki desen ile Penrose Arrowed Rhombi döşemelerinden biri arasında bir benzerlik bulunmuştur.`

Loss: CoSENTLoss with these parameters:

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

Evaluation Dataset

MoritzLaurer/multilingual-nli-26lang-2mil7

Dataset: MoritzLaurer/multilingual-nli-26lang-2mil7 at 510a233
Size: 5,000 evaluation samples
Columns: premise_original, hypothesis_original, score, sentence1, and sentence2

Approximate statistics based on the first 1000 samples:

	premise_original	hypothesis_original	score	sentence1	sentence2
type	string	string	int	string	string
details	min: 5 tokens mean: 30.3 tokens max: 99 tokens	min: 6 tokens mean: 15.11 tokens max: 56 tokens	0: ~34.50% 1: ~29.90% 2: ~35.60%	min: 6 tokens mean: 29.94 tokens max: 106 tokens	min: 5 tokens mean: 15.29 tokens max: 52 tokens

Samples:

premise_original	hypothesis_original	score	sentence1	sentence2
`But the racism charge isn't quirky or wacky--it's demagogy.`	`The accusation of prejudice based on a pedestrian kind of hatred.`	`0`	`Ama ırkçılık suçlaması tuhaf ya da tuhaf değil, bu bir demagoji.`	`Yaya nefretine dayanan önyargı suçlaması.`
`Why would Gates allow the publication of such a book with his byline and photo on the dust jacket?`	`Gates' byline and photo are on the dust jacket`	`0`	`Gates neden böyle bir kitabın basılmasına izin versin ki?`	`Gates'in çizgisi ve fotoğrafı toz ceketin üzerinde.`
`I am a nonsmoker and allergic to cigarette smoke.`	`I do not smoke.`	`0`	`Sigara içmeyen biriyim ve sigara dumanına alerjim var.`	`Sigara içmiyorum.`

Loss: CoSENTLoss with these parameters:

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

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: epoch
per_device_train_batch_size: 32
per_device_eval_batch_size: 64
learning_rate: 2e-05
num_train_epochs: 5
warmup_ratio: 0.1
fp16: True
load_best_model_at_end: True
ddp_find_unused_parameters: False

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: epoch
prediction_loss_only: True
per_device_train_batch_size: 32
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: 2e-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: 5
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: False
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
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
batch_sampler: batch_sampler
multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	loss	tr_ling_spearman_max
0.0320	25	17.17	-	-
0.0639	50	16.4932	-	-
0.0959	75	16.5976	-	-
0.1279	100	15.6991	-	-
0.1598	125	14.876	-	-
0.1918	150	14.4828	-	-
0.2238	175	12.7061	-	-
0.2558	200	10.8687	-	-
0.2877	225	8.3797	-	-
0.3197	250	6.2029	-	-
0.3517	275	5.8228	-	-
0.3836	300	5.811	-	-
0.4156	325	5.8079	-	-
0.4476	350	5.8077	-	-
0.4795	375	5.8035	-	-
0.5115	400	5.8072	-	-
0.5435	425	5.8033	-	-
0.5754	450	5.8086	-	-
0.6074	475	5.81	-	-
0.6394	500	5.7949	-	-
0.6714	525	5.8079	-	-
0.7033	550	5.8057	-	-
0.7353	575	5.8097	-	-
0.7673	600	5.7986	-	-
0.7992	625	5.8051	-	-
0.8312	650	5.8041	-	-
0.8632	675	5.7907	-	-
0.8951	700	5.7991	-	-
0.9271	725	5.8035	-	-
0.9591	750	5.7945	-	-
0.9910	775	5.8077	-	-
1.0	782	-	5.8024	0.0330
1.0230	800	5.6703	-	-
1.0550	825	5.8052	-	-
1.0870	850	5.7936	-	-
1.1189	875	5.7924	-	-
1.1509	900	5.7806	-	-
1.1829	925	5.7835	-	-
1.2148	950	5.7619	-	-
1.2468	975	5.8038	-	-
1.2788	1000	5.779	-	-
1.3107	1025	5.7904	-	-
1.3427	1050	5.7696	-	-
1.3747	1075	5.7919	-	-
1.4066	1100	5.7785	-	-
1.4386	1125	5.7862	-	-
1.4706	1150	5.7703	-	-
1.5026	1175	5.773	-	-
1.5345	1200	5.7627	-	-
1.5665	1225	5.7596	-	-
1.5985	1250	5.7882	-	-
1.6304	1275	5.7828	-	-
1.6624	1300	5.771	-	-
1.6944	1325	5.788	-	-
1.7263	1350	5.7719	-	-
1.7583	1375	5.7846	-	-
1.7903	1400	5.7838	-	-
1.8223	1425	5.7912	-	-
1.8542	1450	5.7686	-	-
1.8862	1475	5.7938	-	-
1.9182	1500	5.7847	-	-
1.9501	1525	5.7952	-	-
1.9821	1550	5.7528	-	-
2.0	1564	-	5.7933	0.0682

Framework Versions

Python: 3.10.12
Sentence Transformers: 3.0.0
Transformers: 4.41.0
PyTorch: 2.3.0+cu121
Accelerate: 0.30.1
Datasets: 2.19.1
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",
}

CoSENTLoss

@online{kexuefm-8847,
    title={CoSENT: A more efficient sentence vector scheme than Sentence-BERT},
    author={Su Jianlin},
    year={2022},
    month={Jan},
    url={https://kexue.fm/archives/8847},
}