metadata

language:
  - bn
  - cs
  - de
  - en
  - et
  - fi
  - fr
  - gu
  - ha
  - hi
  - is
  - ja
  - kk
  - km
  - lt
  - lv
  - pl
  - ps
  - ru
  - ta
  - tr
  - uk
  - xh
  - zh
  - zu
  - ne
  - ro
  - si
tags:
  - sentence-transformers
  - sentence-similarity
  - feature-extraction
  - generated_from_trainer
  - dataset_size:1327190
  - loss:CoSENTLoss
base_model: sentence-transformers/distiluse-base-multilingual-cased-v2
widget:
  - source_sentence: यहाँका केही धार्मिक सम्पदाहरू यस प्रकार रहेका छन्।
    sentences:
      - >-
        A party works journalists from advertisements about a massive Himalayan
        post.
      - Some religious affiliations here remain.
      - >-
        In Spain, the strict opposition of Roman Catholic churches is found to
        have assumed a marriage similar to male beach wives.
  - source_sentence: >-
      "We can use this discovery to target both the assembly and stability of
      the capsid, to either prevent the formation of the virus when it infects
      the host cell, or break it apart after it's formed," Luque said. "This
      could facilitate the characterization and identification of antiviral
      targets for viruses sharing the same icosahedral layout."
    sentences:
      - >-
        FC inter have today released Shefki Kuqi from the club's representative
        team coach duties.
      - >-
        "Wir können diese Entdeckung nutzen, um sowohl die Montage als auch die
        Stabilität des Kapsids anzustreben, um entweder die Bildung des Virus zu
        verhindern, wenn es die Wirtszelle infiziert oder nach seiner Bildung
        auseinanderbricht", sagte Luque. "Dies könnte die Charakterisierung und
        Identifizierung von antiviralen Zielen für Viren erleichtern, die das
        gleiche ikosaedrische Layout teilen".
      - >-
        Quellen sagen, Jones sei „wütend“, als das goldene Mädchen des
        Fernsehens bei einem angespannten Treffen am Dienstag im Hauptquartier
        seines Geschäftsimperiums in Marlow, Buckinghamshire, zugab, dass ihre
        neuen Deals - im Wert von bis zu 1,5 Millionen Pfund - bedeuteten, dass
        sie nicht mehr genug Zeit hatte, sich ihrer Hausbekleidungs- und
        Zubehörmarke Truly zu widmen.
  - source_sentence: >-
      He possesses a pistol with silver bullets for protection from vampires and
      werewolves.
    sentences:
      - >-
        Er besitzt eine Pistole mit silbernen Kugeln zum Schutz vor Vampiren und
        Werwölfen.
      - Bibimbap umfasst Reis, Spinat, Rettich, Bohnensprossen.
      - >-
        BSAC profitierte auch von den großen, aber nicht unbeschränkten
        persönlichen Vermögen von Rhodos und Beit vor ihrem Tod.
  - source_sentence: >-
      To the west of the Badger Head Inlier is the Port Sorell Formation, a
      tectonic mélange of marine sediments and dolerite.
    sentences:
      - >-
        Er brennt einen Speer und brennt Flammen aus seinem Mund, wenn er wütend
        ist.
      - >-
        Westlich des Badger Head Inlier befindet sich die Port Sorell Formation,
        eine tektonische Mischung aus Sedimenten und Dolerit.
      - Public Lynching and Mob Violence under Modi Government
  - source_sentence: >-
      Garnizoana otomană se retrage în sudul Dunării, iar după 164 de ani
      cetatea intră din nou sub stăpânirea europenilor.
    sentences:
      - >-
        This is because, once again, we have taken into account the fact that we
        have adopted a large number of legislative proposals.
      - Helsinki University ranks 75th among universities for 2010.
      - >-
        Ottoman garnisoana is withdrawing into the south of the Danube and,
        after 164 years, it is once again under the control of Europeans.
datasets:
  - RicardoRei/wmt-da-human-evaluation
  - wmt/wmt20_mlqe_task1
pipeline_tag: sentence-similarity
library_name: sentence-transformers
metrics:
  - pearson_cosine
  - spearman_cosine
model-index:
  - name: >-
      SentenceTransformer based on
      sentence-transformers/distiluse-base-multilingual-cased-v2
    results:
      - task:
          type: semantic-similarity
          name: Semantic Similarity
        dataset:
          name: sts eval
          type: sts-eval
        metrics:
          - type: pearson_cosine
            value: 0.42415369784945883
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.4175469519194782
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.0772713008408403
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.13050905562438264
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.16731845692612535
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.18366199919315862
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.3567214608388243
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.3656734148567112
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.41267092498678554
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.41036446071667193
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.5254563854630899
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.4785530551765603
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.31194241573567016
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.2814160300891252
            name: Spearman Cosine
      - task:
          type: semantic-similarity
          name: Semantic Similarity
        dataset:
          name: sts test
          type: sts-test
        metrics:
          - type: pearson_cosine
            value: 0.4253603788235729
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.4166117661445095
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.022187134575214738
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.04647559130832398
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.15979577569463932
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.2074497419832692
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.3698928748443983
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.3757690724227716
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.44937864470538347
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.45866193737582717
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.4466389646053608
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.4158920394678395
            name: Spearman Cosine
          - type: pearson_cosine
            value: 0.33243289478987115
            name: Pearson Cosine
          - type: spearman_cosine
            value: 0.2806845193699054
            name: Spearman Cosine

SentenceTransformer based on sentence-transformers/distiluse-base-multilingual-cased-v2

This is a sentence-transformers model finetuned from sentence-transformers/distiluse-base-multilingual-cased-v2 on the wmt_da, mlqe_en_de, mlqe_en_zh, mlqe_et_en, mlqe_ne_en, mlqe_ro_en and mlqe_si_en datasets. It maps sentences & paragraphs to a 512-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/distiluse-base-multilingual-cased-v2
Maximum Sequence Length: 128 tokens
Output Dimensionality: 512 dimensions
Similarity Function: Cosine Similarity
Training Datasets:
- wmt_da
- mlqe_en_de
- mlqe_en_zh
- mlqe_et_en
- mlqe_ne_en
- mlqe_ro_en
- mlqe_si_en
Languages: bn, cs, de, en, et, fi, fr, gu, ha, hi, is, ja, kk, km, lt, lv, pl, ps, ru, ta, tr, uk, xh, zh, zu, ne, ro, si

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})
  (2): Dense({'in_features': 768, 'out_features': 512, 'bias': True, 'activation_function': 'torch.nn.modules.activation.Tanh'})
)

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("RomainDarous/distiluse-base-multilingual-cased-v2-sts")
# Run inference
sentences = [
    'Garnizoana otomană se retrage în sudul Dunării, iar după 164 de ani cetatea intră din nou sub stăpânirea europenilor.',
    'Ottoman garnisoana is withdrawing into the south of the Danube and, after 164 years, it is once again under the control of Europeans.',
    'This is because, once again, we have taken into account the fact that we have adopted a large number of legislative proposals.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 512]

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

Evaluation

Metrics

Semantic Similarity

Datasets: sts-eval, sts-test, sts-test, sts-test, sts-test, sts-test, sts-test and sts-test
Evaluated with EmbeddingSimilarityEvaluator

Metric	sts-eval	sts-test
pearson_cosine	0.4242	0.3324
spearman_cosine	0.4175	0.2807

Semantic Similarity

Dataset: sts-eval
Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.0773
spearman_cosine	0.1305

Semantic Similarity

Dataset: sts-eval
Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.1673
spearman_cosine	0.1837

Semantic Similarity

Dataset: sts-eval
Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.3567
spearman_cosine	0.3657

Semantic Similarity

Dataset: sts-eval
Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.4127
spearman_cosine	0.4104

Semantic Similarity

Dataset: sts-eval
Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.5255
spearman_cosine	0.4786

Semantic Similarity

Dataset: sts-eval
Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.3119
spearman_cosine	0.2814

Training Details

Training Datasets

wmt_da

Dataset: wmt_da at 301de38
Size: 1,285,190 training samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 4 tokens
mean: 37.09 tokens
max: 128 tokens

min: 4 tokens
mean: 37.12 tokens
max: 128 tokens

min: 0.0
mean: 0.7
max: 1.0

	sentence1	sentence2	score
type	string	string	float
details	min: 4 tokens mean: 37.09 tokens max: 128 tokens	min: 4 tokens mean: 37.12 tokens max: 128 tokens	min: 0.0 mean: 0.7 max: 1.0

Samples:

sentence1	sentence2	score
`Z dat ÚZIS také vyplývá, že se zastavil úbytek zdravotních sester v nemocnicích.`	`The data from the IHIS also shows that the decline of nurses in hospitals has stopped.`	`0.47`
`Я был самым гордым, самым пьяным девственником, которого кто-либо когда-либо видел.`	`I was the proudest, most drunk virgin anyone had ever seen.`	`0.99`
`Das Trampolinspringen hat einen gewissen Außenseitercharme, teilweise weil es für das unaufgeklärte Ohr passender für eine Clownsschule als die für die Olympischen Spiele klingt.`	`The trampoline jumping has some outsider charm, in part because it sounds more appropriate for the unenlightened ear for a clowns school than the one for the Olympics.`	`0.81`

Loss: CoSENTLoss with these parameters:

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

mlqe_en_de

Dataset: mlqe_en_de at 0783ed2
Size: 7,000 training samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 11 tokens
mean: 23.78 tokens
max: 44 tokens

min: 11 tokens
mean: 26.51 tokens
max: 54 tokens

min: 0.06
mean: 0.86
max: 1.0

	sentence1	sentence2	score
type	string	string	float
details	min: 11 tokens mean: 23.78 tokens max: 44 tokens	min: 11 tokens mean: 26.51 tokens max: 54 tokens	min: 0.06 mean: 0.86 max: 1.0

Samples:

sentence1	sentence2	score
`Early Muslim traders and merchants visited Bengal while traversing the Silk Road in the first millennium.`	`Frühe muslimische Händler und Kaufleute besuchten Bengalen, während sie im ersten Jahrtausend die Seidenstraße durchquerten.`	`0.9233333468437195`
`While Fran dissipated shortly after that, the tropical wave progressed into the northeastern Pacific Ocean.`	`Während Fran kurz danach zerstreute, entwickelte sich die tropische Welle in den nordöstlichen Pazifischen Ozean.`	`0.8899999856948853`
`Distressed securities include such events as restructurings, recapitalizations, and bankruptcies.`	`Zu den belasteten Wertpapieren gehören Restrukturierungen, Rekapitalisierungen und Insolvenzen.`	`0.9300000071525574`

Loss: CoSENTLoss with these parameters:

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

mlqe_en_zh

Dataset: mlqe_en_zh at 0783ed2
Size: 7,000 training samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 9 tokens
mean: 24.09 tokens
max: 47 tokens

min: 12 tokens
mean: 29.93 tokens
max: 74 tokens

min: 0.01
mean: 0.68
max: 0.98

	sentence1	sentence2	score
type	string	string	float
details	min: 9 tokens mean: 24.09 tokens max: 47 tokens	min: 12 tokens mean: 29.93 tokens max: 74 tokens	min: 0.01 mean: 0.68 max: 0.98

Samples:

sentence1	sentence2	score
`In the late 1980s, the hotel's reputation declined, and it functioned partly as a "backpackers hangout."`	`在 20 世纪 80 年代末，这家旅馆的声誉下降了，部分地起到了 "背包吊销" 的作用。`	`0.40666666626930237`
`From 1870 to 1915, 36 million Europeans migrated away from Europe.`	`从 1870 年到 1915 年， 3, 600 万欧洲人从欧洲移民。`	`0.8333333730697632`
`In some photos, the footpads did press into the regolith, especially when they moved sideways at touchdown.`	`在一些照片中，脚垫确实挤进了后台，尤其是当他们在触地时侧面移动时。`	`0.33000001311302185`

Loss: CoSENTLoss with these parameters:

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

mlqe_et_en

Dataset: mlqe_et_en at 0783ed2
Size: 7,000 training samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 14 tokens
mean: 31.88 tokens
max: 63 tokens

min: 11 tokens
mean: 24.57 tokens
max: 56 tokens

min: 0.03
mean: 0.67
max: 1.0

	sentence1	sentence2	score
type	string	string	float
details	min: 14 tokens mean: 31.88 tokens max: 63 tokens	min: 11 tokens mean: 24.57 tokens max: 56 tokens	min: 0.03 mean: 0.67 max: 1.0

Samples:

sentence1	sentence2	score
`Gruusias vahistati president Mihhail Saakašvili pressibüroo nõunik Simon Kiladze, keda süüdistati spioneerimises.`	`In Georgia, an adviser to the press office of President Mikhail Saakashvili, Simon Kiladze, was arrested and accused of spying.`	`0.9466666579246521`
`Nii teadmissotsioloogia pooldajad tavaliselt Kuhni tõlgendavadki, arendades tema vaated sõnaselgeks relativismiks.`	`This is how supporters of knowledge sociology usually interpret Kuhn by developing his views into an explicit relativism.`	`0.9366666674613953`
`18. jaanuaril 2003 haarasid mitmeid Canberra eeslinnu võsapõlengud, milles hukkus neli ja sai vigastada 435 inimest.`	`On 18 January 2003, several of the suburbs of Canberra were seized by debt fires which killed four people and injured 435 people.`	`0.8666666150093079`

Loss: CoSENTLoss with these parameters:

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

mlqe_ne_en

Dataset: mlqe_ne_en at 0783ed2
Size: 7,000 training samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 17 tokens
mean: 40.67 tokens
max: 77 tokens

min: 9 tokens
mean: 24.66 tokens
max: 128 tokens

min: 0.01
mean: 0.39
max: 1.0

	sentence1	sentence2	score
type	string	string	float
details	min: 17 tokens mean: 40.67 tokens max: 77 tokens	min: 9 tokens mean: 24.66 tokens max: 128 tokens	min: 0.01 mean: 0.39 max: 1.0

Samples:

sentence1	sentence2	score
`सामान्‍य बजट प्रायः फेब्रुअरीका अंतिम कार्य दिवसमा लाईन्छ।`	`A normal budget is usually awarded to the digital working day of February.`	`0.5600000023841858`
`कविताका यस्ता स्वरूपमा दुई, तिन वा चार पाउसम्मका मुक्तक, हाइकु, सायरी र लोकसूक्तिहरू पर्दछन् ।`	`The book consists of two, free of her or four paulets, haiku, Sairi, and locus in such forms.`	`0.23666666448116302`
`ब्रिट्नीले यस बारेमा प्रतिक्रिया ब्यक्ता गरदै भनिन,"कुन ठूलो कुरा हो र?`	`Britney did not respond to this, saying "which is a big thing and a big thing?`	`0.21666665375232697`

Loss: CoSENTLoss with these parameters:

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

mlqe_ro_en

Dataset: mlqe_ro_en at 0783ed2
Size: 7,000 training samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 12 tokens
mean: 29.44 tokens
max: 60 tokens

min: 10 tokens
mean: 22.38 tokens
max: 65 tokens

min: 0.01
mean: 0.68
max: 1.0

	sentence1	sentence2	score
type	string	string	float
details	min: 12 tokens mean: 29.44 tokens max: 60 tokens	min: 10 tokens mean: 22.38 tokens max: 65 tokens	min: 0.01 mean: 0.68 max: 1.0

Samples:

sentence1	sentence2	score
`Orașul va fi împărțit în patru districte, iar suburbiile în 10 mahalale.`	`The city will be divided into four districts and suburbs into 10 mahalals.`	`0.4699999988079071`
`La scurt timp după aceasta, au devenit cunoscute debarcările germane de la Trondheim, Bergen și Stavanger, precum și luptele din Oslofjord.`	`In the light of the above, the Authority concludes that the aid granted to ADIF is compatible with the internal market pursuant to Article 61 (3) (c) of the EEA Agreement.`	`0.02666666731238365`
`Până în vara 1791, în Clubul iacobinilor au dominat reprezentanții monarhismului liberal constituțional.`	`Until the summer of 1791, representatives of liberal constitutional monarchism dominated in the Jacobins Club.`	`0.8733333349227905`

Loss: CoSENTLoss with these parameters:

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

mlqe_si_en

Dataset: mlqe_si_en at 0783ed2
Size: 7,000 training samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 8 tokens
mean: 18.19 tokens
max: 38 tokens

min: 9 tokens
mean: 22.31 tokens
max: 128 tokens

min: 0.01
mean: 0.51
max: 1.0

	sentence1	sentence2	score
type	string	string	float
details	min: 8 tokens mean: 18.19 tokens max: 38 tokens	min: 9 tokens mean: 22.31 tokens max: 128 tokens	min: 0.01 mean: 0.51 max: 1.0

Samples:

sentence1	sentence2	score
`ඇපලෝ 4 සැටර්න් V බූස්ටරයේ ප්‍රථම පර්යේෂණ පියාසැරිය විය.`	`The first research flight of the Apollo 4 Saturn V Booster.`	`0.7966666221618652`
`මෙහි අවපාතය සැලකීමේ දී, මෙහි 48%ක අවරෝහණය $ මිලියන 125කට අධික චිත්‍රපටයක් ලද තෙවන කුඩාම අවපාතය වේ.`	`In conjunction with the depression here, 48 % of obesity here is the third smallest depression in over $ 125 million film.`	`0.17666666209697723`
`එසේම "බකමූණන් මගින් මෙම රාක්ෂසියගේ රාත්‍රී හැසිරීම සංකේතවත් වන බව" පවසයි.`	`Also "the owl says that this monster's night behavior is symbolic".`	`0.8799999952316284`

Loss: CoSENTLoss with these parameters:

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

Evaluation Datasets

wmt_da

Dataset: wmt_da at 301de38
Size: 1,285,190 evaluation samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 4 tokens
mean: 36.52 tokens
max: 128 tokens

min: 4 tokens
mean: 36.59 tokens
max: 128 tokens

min: 0.0
mean: 0.7
max: 1.0

	sentence1	sentence2	score
type	string	string	float
details	min: 4 tokens mean: 36.52 tokens max: 128 tokens	min: 4 tokens mean: 36.59 tokens max: 128 tokens	min: 0.0 mean: 0.7 max: 1.0

Samples:

sentence1	sentence2	score
`The note adds that should the departure from the White House be delayed, a second aircrew would be needed for the return flight due to duty-hour restrictions.`	`V poznámce se dodává, že pokud by se odlet z Bílého domu zpozdil, byla by pro zpáteční let kvůli omezení pracovní doby nutná druhá letecká posádka.`	`0.95`
`上半年电信网络诈骗犯罪上升七成最高检总结特点-中新网`	`In the first half of the year, telecommunication network fraud crimes rose by 70%. The highest inspection summary characteristics-Zhongxin.com`	`0.72`
`Als zentrale Herausforderungen für den Bundesnachrichtendienst (BND) nannte Merkel den Kampf gegen die Verbreitung von Falschmeldungen im Internet und die Abwehr von Cyberattacken.`	`Merkel a cité la lutte contre la propagation de fausses nouvelles en ligne et la défense contre les cyberattaques comme des défis majeurs pour le service fédéral de renseignement (BND).`	`0.87`

Loss: CoSENTLoss with these parameters:

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

mlqe_en_de

Dataset: mlqe_en_de at 0783ed2
Size: 1,000 evaluation samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 11 tokens
mean: 24.11 tokens
max: 49 tokens

min: 11 tokens
mean: 26.66 tokens
max: 55 tokens

min: 0.03
mean: 0.81
max: 1.0

	sentence1	sentence2	score
type	string	string	float
details	min: 11 tokens mean: 24.11 tokens max: 49 tokens	min: 11 tokens mean: 26.66 tokens max: 55 tokens	min: 0.03 mean: 0.81 max: 1.0

Samples:

sentence1	sentence2	score
`Resuming her patrols, Constitution managed to recapture the American sloop Neutrality on 27 March and, a few days later, the French ship Carteret.`	`Mit der Wiederaufnahme ihrer Patrouillen gelang es der Verfassung, am 27. März die amerikanische Schleuderneutralität und wenige Tage später das französische Schiff Carteret zurückzuerobern.`	`0.9033333659172058`
`Blaine's nomination alienated many Republicans who viewed Blaine as ambitious and immoral.`	`Blaines Nominierung entfremdete viele Republikaner, die Blaine als ehrgeizig und unmoralisch betrachteten.`	`0.9216666221618652`
`This initiated a brief correspondence between the two which quickly descended into political rancor.`	`Dies leitete eine kurze Korrespondenz zwischen den beiden ein, die schnell zu politischem Groll abstieg.`	`0.878333330154419`

Loss: CoSENTLoss with these parameters:

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

mlqe_en_zh

Dataset: mlqe_en_zh at 0783ed2
Size: 1,000 evaluation samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 9 tokens
mean: 23.75 tokens
max: 49 tokens

min: 11 tokens
mean: 29.56 tokens
max: 67 tokens

min: 0.26
mean: 0.65
max: 0.9

	sentence1	sentence2	score
type	string	string	float
details	min: 9 tokens mean: 23.75 tokens max: 49 tokens	min: 11 tokens mean: 29.56 tokens max: 67 tokens	min: 0.26 mean: 0.65 max: 0.9

Samples:

sentence1	sentence2	score
`Freeman briefly stayed with the king before returning to Accra via Whydah, Ahgwey and Little Popo.`	`弗里曼在经过惠达、阿格威和小波波回到阿克拉之前与国王一起住了一会儿。`	`0.6683333516120911`
`Fantastic Fiction "Scratches in the Sky, Ben Peek, Agog!`	`奇特的虚构 "天空中的碎片，本佩克，阿戈！`	`0.71833336353302`
`For Hermann Keller, the running quavers and semiquavers "suffuse the setting with health and strength."`	`对赫尔曼 · 凯勒来说，跑步的跳跃者和半跳跃者 "让环境充满健康和力量" 。`	`0.7066666483879089`

Loss: CoSENTLoss with these parameters:

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

mlqe_et_en

Dataset: mlqe_et_en at 0783ed2
Size: 1,000 evaluation samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 12 tokens
mean: 32.4 tokens
max: 58 tokens

min: 10 tokens
mean: 24.87 tokens
max: 47 tokens

min: 0.03
mean: 0.6
max: 0.99

	sentence1	sentence2	score
type	string	string	float
details	min: 12 tokens mean: 32.4 tokens max: 58 tokens	min: 10 tokens mean: 24.87 tokens max: 47 tokens	min: 0.03 mean: 0.6 max: 0.99

Samples:

sentence1	sentence2	score
`Jackson pidas seal kõne, öeldes, et James Brown on tema suurim inspiratsioon.`	`Jackson gave a speech there saying that James Brown is his greatest inspiration.`	`0.9833333492279053`
`Kaanelugu rääkis loo kolme ungarlase üleelamistest Ungari revolutsiooni päevil.`	`The life of the Man spoke of a story of three Hungarians living in the days of the Hungarian Revolution.`	`0.28999999165534973`
`Teise maailmasõja ajal oli ta mitme Saksa juhatusele alluvate eesti väeosa ülem.`	`During World War II, he was the commander of several of the German leadership.`	`0.4516666829586029`

Loss: CoSENTLoss with these parameters:

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

mlqe_ne_en

Dataset: mlqe_ne_en at 0783ed2
Size: 1,000 evaluation samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 17 tokens
mean: 41.03 tokens
max: 85 tokens

min: 10 tokens
mean: 24.77 tokens
max: 128 tokens

min: 0.05
mean: 0.36
max: 0.92

	sentence1	sentence2	score
type	string	string	float
details	min: 17 tokens mean: 41.03 tokens max: 85 tokens	min: 10 tokens mean: 24.77 tokens max: 128 tokens	min: 0.05 mean: 0.36 max: 0.92

Samples:

sentence1	sentence2	score
`१८९२ तिर भवानीदत्त पाण्डेले 'मुद्रा राक्षस'को अनुवाद गरे।`	`Around 1892, Bhavani Pandit translated the 'money monster'.`	`0.8416666388511658`
`यस बच्चाको मुखले आमाको स्तन यस बच्चाको मुखले आमाको स्तन राम्ररी च्यापेको छ ।`	`The breasts of this child's mouth are taped well with the mother's mouth.`	`0.2150000035762787`
`बुवाको बन्दुक चोरेर हिँडेका बराललाई केआई सिंहले अब गोली ल्याउन लगाए ।...`	`Kei Singh, who stole the boy's closet, took the bullet to bring it now..`	`0.27000001072883606`

Loss: CoSENTLoss with these parameters:

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

mlqe_ro_en

Dataset: mlqe_ro_en at 0783ed2
Size: 1,000 evaluation samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 14 tokens
mean: 30.25 tokens
max: 59 tokens

min: 6 tokens
mean: 22.7 tokens
max: 55 tokens

min: 0.01
mean: 0.68
max: 1.0

	sentence1	sentence2	score
type	string	string	float
details	min: 14 tokens mean: 30.25 tokens max: 59 tokens	min: 6 tokens mean: 22.7 tokens max: 55 tokens	min: 0.01 mean: 0.68 max: 1.0

Samples:

sentence1	sentence2	score
`Cornwallis se afla înconjurat pe uscat de forțe armate net superioare și retragerea pe mare era îndoielnică din cauza flotei franceze.`	`Cornwallis was surrounded by shore by higher armed forces and the sea withdrawal was doubtful due to the French fleet.`	`0.8199999928474426`
`thumbrightuprightDansatori [[cretani de muzică tradițională.`	`Number of employees employed in the production of the like product in the Union.`	`0.009999999776482582`
`Potrivit documentelor vremii și tradiției orale, aceasta a fost cea mai grea perioadă din istoria orașului.`	`According to the documents of the oral weather and tradition, this was the hardest period in the city's history.`	`0.5383332967758179`

Loss: CoSENTLoss with these parameters:

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

mlqe_si_en

Dataset: mlqe_si_en at 0783ed2
Size: 1,000 evaluation samples
Columns: sentence1, sentence2, and score
Approximate statistics based on the first 1000 samples:
sentence1 sentence2 score
type string string float
details
min: 8 tokens
mean: 18.12 tokens
max: 36 tokens

min: 7 tokens
mean: 22.18 tokens
max: 128 tokens

min: 0.03
mean: 0.51
max: 0.99

	sentence1	sentence2	score
type	string	string	float
details	min: 8 tokens mean: 18.12 tokens max: 36 tokens	min: 7 tokens mean: 22.18 tokens max: 128 tokens	min: 0.03 mean: 0.51 max: 0.99

Samples:

sentence1	sentence2	score
`එයට ශි්‍ර ලංකාවේ සාමය ඇති කිරිමටත් නැති කිරිමටත් පුළුවන්.`	`It can also cause peace in Sri Lanka.`	`0.3199999928474426`
`ඔහු මනෝ විද්‍යාව, සමාජ විද්‍යාව, ඉතිහාසය හා සන්නිවේදනය යන විෂය ක්ෂේත්‍රයන් පිලිබදවද අධ්‍යයනයන් සිදු කිරීමට උත්සාහ කරන ලදි.`	`He attempted to do subjects in psychology, sociology, history and communication.`	`0.5366666913032532`
`එහෙත් කිසිදු මිනිසෙක්‌ හෝ ගැහැනියෙක්‌ එලිමහනක නොවූහ.`	`But no man or woman was eliminated.`	`0.2783333361148834`

Loss: CoSENTLoss with these parameters:

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

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 64
per_device_eval_batch_size: 64
num_train_epochs: 2
warmup_ratio: 0.1

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: steps
prediction_loss_only: True
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
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: 2
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: False
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: 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: None
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: batch_sampler
multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	wmt da loss	mlqe en de loss	mlqe en zh loss	mlqe et en loss	mlqe ne en loss	mlqe ro en loss	mlqe si en loss	sts-eval_spearman_cosine	sts-test_spearman_cosine
0.4	6690	7.8421	7.5547	7.5619	7.5555	7.5327	7.5354	7.5109	7.5564	0.1989	-
0.8	13380	7.552	7.5420	7.5757	7.5739	7.5185	7.5126	7.4994	7.5511	0.2336	-
1.2	20070	7.5216	7.5465	7.6072	7.5942	7.5217	7.5141	7.4871	7.5471	0.2694	-
1.6	26760	7.5024	7.5329	7.6123	7.5814	7.5230	7.5141	7.4679	7.5379	0.2866	-
2.0	33450	7.495	7.5252	7.6106	7.5756	7.5201	7.5128	7.4725	7.5417	0.2814	0.2807

Framework Versions

Python: 3.11.10
Sentence Transformers: 3.3.1
Transformers: 4.47.1
PyTorch: 2.3.1+cu121
Accelerate: 1.2.1
Datasets: 3.2.0
Tokenizers: 0.21.0

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},
}