bobox/DeBERTaV3-small-SenTra-AdaptiveLayers-AllSoft-HighTemp

SentenceTransformer based on microsoft/deberta-v3-small

This is a sentence-transformers model finetuned from microsoft/deberta-v3-small on the nli-pairs, sts-label, vitaminc-pairs, qnli-contrastive, scitail-pairs-qa, scitail-pairs-pos, xsum-pairs, compression-pairs, sciq_pairs, qasc_pairs, openbookqa_pairs, msmarco_pairs, nq_pairs, trivia_pairs, quora_pairs and gooaq_pairs 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. \n\n train_loss = AdaptiveLayerLoss(model=model, loss=train_loss, n_layers_per_step = -1, last_layer_weight = 1.5, prior_layers_weight= 0.15, kl_div_weight = 2, kl_temperature= 2,) num_epochs = 4, learning_rate = 2e-5, warmup_ratio=0.25, weight_decay = 5e-7, schedule = "cosine_with_restarts", num_cycles = 5

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: microsoft/deberta-v3-small
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 tokens
• Similarity Function: Cosine Similarity
• Training Datasets:
  • nli-pairs
  • sts-label
  • vitaminc-pairs
  • qnli-contrastive
  • scitail-pairs-qa
  • scitail-pairs-pos
  • xsum-pairs
  • compression-pairs
  • sciq_pairs
  • qasc_pairs
  • openbookqa_pairs
  • msmarco_pairs
  • nq_pairs
  • trivia_pairs
  • quora_pairs
  • gooaq_pairs
• 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: DebertaV2Model 
  (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("bobox/DeBERTaV3-small-SenTra-AdaptiveLayers-AllSoft-HighTemp")
# Run inference
sentences = [
    'How many hymns of Luther were included in the Achtliederbuch?',
    "Luther's hymns were included in early Lutheran hymnals and spread the ideas of the Reformation.",
    'the ABC News building was renamed Peter Jennings Way in 2006 in honor of the recently deceased longtime ABC News chief anchor and anchor of World News Tonight.',
]
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

Semantic Similarity

• Dataset: sts-test
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.5667
spearman_cosine	0.5551
pearson_manhattan	0.5771
spearman_manhattan	0.5724
pearson_euclidean	0.5619
spearman_euclidean	0.5593
pearson_dot	0.2353
spearman_dot	0.2422
pearson_max	0.5771
spearman_max	0.5724

Semantic Similarity

• Dataset: sts-test
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.5667
spearman_cosine	0.5551
pearson_manhattan	0.5771
spearman_manhattan	0.5724
pearson_euclidean	0.5619
spearman_euclidean	0.5593
pearson_dot	0.2353
spearman_dot	0.2422
pearson_max	0.5771
spearman_max	0.5724

Semantic Similarity

• Dataset: sts-test
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.5667
spearman_cosine	0.5551
pearson_manhattan	0.5771
spearman_manhattan	0.5724
pearson_euclidean	0.5619
spearman_euclidean	0.5593
pearson_dot	0.2353
spearman_dot	0.2422
pearson_max	0.5771
spearman_max	0.5724

Semantic Similarity

• Dataset: sts-test
• Evaluated with EmbeddingSimilarityEvaluator

Metric	Value
pearson_cosine	0.5667
spearman_cosine	0.5551
pearson_manhattan	0.5771
spearman_manhattan	0.5724
pearson_euclidean	0.5619
spearman_euclidean	0.5593
pearson_dot	0.2353
spearman_dot	0.2422
pearson_max	0.5771
spearman_max	0.5724

Training Details

Training Datasets

nli-pairs

• Dataset: nli-pairs at d482672
• Size: 6,500 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 5 tokens mean: 16.62 tokens max: 62 tokens	min: 4 tokens mean: 9.46 tokens max: 29 tokens

• Samples:

  sentence1	sentence2
  A person on a horse jumps over a broken down airplane.	A person is outdoors, on a horse.
  Children smiling and waving at camera	There are children present
  A boy is jumping on skateboard in the middle of a red bridge.	The boy does a skateboarding trick.

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

sts-label

• Dataset: sts-label at ab7a5ac
• Size: 5,749 training samples
• Columns: sentence1, sentence2, and score
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	score
  type	string	string	float
  details	min: 6 tokens mean: 9.81 tokens max: 27 tokens	min: 5 tokens mean: 9.74 tokens max: 25 tokens	min: 0.0 mean: 0.54 max: 1.0

• Samples:

  sentence1	sentence2	score
  A plane is taking off.	An air plane is taking off.	1.0
  A man is playing a large flute.	A man is playing a flute.	0.76
  A man is spreading shreded cheese on a pizza.	A man is spreading shredded cheese on an uncooked pizza.	0.76

• Loss: CoSENTLoss with these parameters:

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

vitaminc-pairs

• Dataset: vitaminc-pairs at be6febb
• Size: 3,194 training samples
• Columns: label, sentence1, and sentence2
• Approximate statistics based on the first 1000 samples:

  	label	sentence1	sentence2
  type	int	string	string
  details	1: 100.00%	min: 6 tokens mean: 15.8 tokens max: 75 tokens	min: 8 tokens mean: 38.29 tokens max: 512 tokens

• Samples:

  label	sentence1	sentence2
  1	Kyle Kendricks was otherwise called the Professor .	Chicago Cubs ( �present ) } } Kyle Christian Hendricks ( born December 7 , 1989 ) , nicknamed '' The Proffessor , '' '' is an American professional baseball pitcher for the Chicago Cubs of Major League Baseball ( MLB ) . ''
  1	Since 1982 , 533 people have been executed in Texas .	Since the death penalty was re-instituted in the United States with the 1976 Gregg v. Georgia decision , Texas has executed more inmates than any other state , beginning in 1982 with the execution of Charles Brooks , Jr.. Since 1982 , 533 people have been executed in Texas. 1923 , the Texas Department of Criminal Justice ( TDCJ ) has been in charge of executions in the state .
  1	Hilltop Hoods have released two `` restrung '' albums .	The group released its first extended play , Back Once Again , in 1997 and have subsequently released seven studio albums , two '' restrung '' '' albums and three DVDs . ''

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

qnli-contrastive

• Dataset: qnli-contrastive at bcdcba7
• Size: 4,000 training samples
• Columns: sentence1, sentence2, and label
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	label
  type	string	string	int
  details	min: 6 tokens mean: 13.79 tokens max: 40 tokens	min: 6 tokens mean: 35.8 tokens max: 499 tokens	0: 100.00%

• Samples:

  sentence1	sentence2	label
  Vinters have adopted solar technology to do what?	More recently the technology has been embraced by vinters, who use the energy generated by solar panels to power grape presses.	0
  Who did Madonna's look and style of dressing influence?	It attracted the attention of organizations who complained that the song and its accompanying video promoted premarital sex and undermined family values, and moralists sought to have the song and video banned.	0
  In addition to hearing him play, what else did people seek from Chopin in London?	The Prince, who was himself a talented musician, moved close to the keyboard to view Chopin's technique.	0

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "OnlineContrastiveLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

scitail-pairs-qa

• Dataset: scitail-pairs-qa at 0cc4353
• Size: 4,300 training samples
• Columns: sentence2 and sentence1
• Approximate statistics based on the first 1000 samples:

  	sentence2	sentence1
  type	string	string
  details	min: 7 tokens mean: 16.0 tokens max: 41 tokens	min: 7 tokens mean: 14.71 tokens max: 34 tokens

• Samples:

  sentence2	sentence1
  The fetal period lasts approximately 30 weeks weeks.	Approximately how many weeks does the fetal period last?
  Corals build hard exoskeletons that grow to become coral reefs.	Corals build hard exoskeletons that grow to become what?
  A voltaic cell generates an electric current through a reaction known as a(n) spontaneous redox.	A voltaic cell uses what type of reaction to generate an electric current

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

scitail-pairs-pos

• Dataset: scitail-pairs-pos at 0cc4353
• Size: 2,200 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 7 tokens mean: 23.76 tokens max: 74 tokens	min: 7 tokens mean: 15.27 tokens max: 41 tokens

• Samples:

  sentence1	sentence2
  As the water vapor cools, it condenses , forming tiny droplets in clouds.	Clouds are formed from water droplets.
  Poison ivy is green, with three leaflets on each leaf, grows as a shrub or vine, and may be in your yard.	Poison ivy typically has three groups of leaves.
  (Formic acid is the poison found in the > sting of fire ants.)	Formic acid is found in the secretions of stinging ants.

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

xsum-pairs

• Dataset: xsum-pairs at 788ddaf
• Size: 2,500 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 14 tokens mean: 345.33 tokens max: 512 tokens	min: 6 tokens mean: 27.11 tokens max: 60 tokens

• Samples:

  sentence1	sentence2
  Rahim Kalantar told the BBC his son Ali, 18, travelled to Syria with two friends from Coventry in March and believed he was now fighting with Isis. He said he was sent "down this road" by an imam - who denied the allegations. Up to 500 Britons are thought to have travelled to the Middle East to fight in the conflict, officials say. Mr Kalantar - speaking to BBC Two's Newsnight, in collaboration with the BBC's Afghan Service and Newsday - said he worries about his son Ali "every minute" and that his grief is "limitless". He said he believed Ali - who was planning to study computer science at university - had been radicalised during classes at a mosque after evening prayer. "He [the imam] encouraged them and sent them down this road," he said. The BBC contacted the mosque to speak to the imam, who refused to give an interview but said he completely denied the allegations. Ali is believed to have travelled to Syria with Rashed Amani, also 18, who had been studying business at Coventry University. Rashed's father, Khabir, said family members had travelled to the Turkish-Syrian border in the hope of finding the boys, but came back "empty-handed" after searching for more than two weeks. He said he did not know what had happened to his son, who he fears has joined Isis - the militant-led group that has made rapid advances through Iraq in recent weeks. "Maybe somebody worked with him, I don't know. Maybe somebody brainwashed him because he was not like that," he said. The third teenager, Moh Ismael, is also believed to be in Syria with his friends. He is understood to have posted a message on Twitter saying he was with Isis. It comes after Britons - including Reyaad Khan and Nasser Muthana from Cardiff - featured in an apparent recruitment video for jihadists in Iraq and Syria. The video was posted online on Friday by accounts with links to Isis. The BBC has learned a third Briton in the video is from Aberdeen. The man, named locally as Raqib, grew up in Scotland but was originally from Bangladesh. Lord Carlile, a former independent reviewer of terrorism laws, told the BBC that the Muslim community was best placed to stop jihadists recruiting in the UK. The Liberal Democrat peer also said the UK needed to reintroduce tougher measures to stop terrorism. It comes after former MI6 director, Richard Barrett, said security services would not be able to track all Britons who return to the UK after fighting in Syria. He said the number of those posing a threat would be small but unpredictable. The Metropolitan Police has insisted it has the tools to monitor British jihadists returning from that country. Shiraz Maher, a radicalisation expert, told Newsnight that social media was now acting as a recruitment ground for potential jihadists in the UK. "You have hundreds of foreign fighters on the ground who in real time are giving you a live feed of what is happening and they are engaged in a conversation. "It is these individual people who have been empowered to become recruiters in their own right," he said. Lord Carlile said the "most important partners" in preventing young Muslims from being radicalised were the "Muslim communities themselves". "Mothers, wives, sisters do not want their husbands, brothers, sons to become valid jihadists and run the risk of being killed in a civil war," he told the programme. He also told BBC Radio 4's World at One programme that the government should look at reintroducing "something like control orders", which were scrapped in 2011 and replaced with the less restrictive Terrorism Prevention and Investigation Measures (TPims). He said: "We need to look at preventing violent extremism before people leave the country and also we need to look for further measures."	The father of a British teenager who travelled to Syria to join jihadists believes his son was radicalised by an imam at a UK mosque.
  Jawad Fairooz and Matar Matar were detained in May after resigning from parliament in protest at the handling of the protests. Mr Matar told the BBC they had been tortured in prison. They were prosecuted in a security court on charges of taking part in illegal protests and defaming the country. It is not clear if they still face trial in a civilian court. Civilian courts took over jurisdiction after King Hamad Bin Issa Al Khalifa lifted a state of emergency in June. Mr Matar told the BBC he believed his arrest had been intended to put a pressure on his al-Wifaq party. "At some stages we were tortured," he said. "In one of the cases we were beaten." Human rights lawyer Mohamed al-Tajir was also released. He was detained in April having defended people arrested during the Saudi-backed suppression of protests in March. Correspondents say their release appears to be an attempt at defusing tensions in the country, a key US ally in the region that hosts the US Navy's 5th Fleet. Bahrain's King Hamad Bin Issa Al Khalifa recently accepted a series of reforms drawn up by a government-backed committee created to address grievances that emerged during the protests. The kingdom's Shia community makes up about 70% of the population but many say they are discriminated against by the minority Sunni monarchy.	Bahrain has freed two former Shia opposition MPs arrested in the wake of widespread anti-government protests.
  Liverpool City Region, in case you were wondering, includes Merseyside's five councils (Knowsley, Liverpool, Sefton, St Helens, and Wirral) as well as Halton in Cheshire. Who are the eight candidates desperate for your support on 4 May, though, and what are their priorities? BBC Radio Merseyside's political reporter Claire Hamilton has produced a potted biography for each of them. We're also asking all of them for a "minute manifesto" video. Candidates are listed below in alphabetical order Roger Bannister, Trade Union & Socialist Coalition Veteran trade unionist Roger Bannister believes the Liverpool City Region Combined Authority should never have approved the contract for a fleet of new driver-only Merseyrail trains. He says he would seek to reverse this decision. He also believes local authorities have passed harmful austerity budgets on people struggling to make ends meet. He stood for Liverpool city mayor in 2016, coming fourth with 5% of the vote. Paul Breen, Get the Coppers off the Jury Paul Breen is a resident of Norris Green, Liverpool and became the last candidate to be nominated. He is listed as treasurer of the party on the Electoral Commission's website, with Patricia Breen listed as deputy treasurer. He has not yet released any material detailing his manifesto but told the BBC the title of his campaign speaks for itself. He simply does not believe that police officers should be allowed to serve on juries. Mr Breen declined to provide a "minute manifesto" Tony Caldeira, Conservative Born in Liverpool and educated in St Helens, Tony Caldeira started out working on a stall selling cushions made by his mother at Liverpool's Great Homer Street market. His business expanded and now operates in Kirkby, distributing world-wide. Mr Caldeira has stood for Liverpool mayor twice, coming sixth in 2016 with just under 4% of the vote. He has pledged to improve the area's transport network, speed up the planning process and build homes and workplaces on brownfield sites rather than green spaces. Carl Cashman, Liberal Democrats Born in Whiston, Knowsley, Carl Cashman is leader of the Liberal Democrat group on Knowsley Council. He and his two Lib Dem council colleagues were elected in 2016, breaking a four-year period when Labour was the only party represented. Aged 25, he's the youngest of the candidates. Mr Cashman believes maintaining strong ties with Europe and the region will be key, and has pledged to open a Liverpool City Region embassy in Brussels. He also wants to better integrate ticketing across public transport and make the current Walrus card more similar to the Oyster card used by Londoners. Tom Crone, Green Party Tom Crone is leader of the Green group on Liverpool City Council. He won 10% of the vote in the mayoral elections in Liverpool in 2016 and came third. Originally from Norwich, he has lived in Liverpool since 2000 after arriving as a student. Mr Crone is keen to see a shift away from traditional heavy industry in the city region towards greener "tech" industries. He's also passionate about making public transport more affordable and environmentally friendly. He says he'll look to prioritise new routes for cyclists and pedestrians. Tabitha Morton, Women's Equality Party Tabitha Morton was born in Netherton, Sefton. She left school with no formal qualifications, and started work at 16 at a local market, and later in cleaning. She was taken on for NVQ training by a company in Liverpool, and stayed on to train others. She now works for a global manufacturer, in what she describes as "a male-dominated industry". She says she would prioritise grants for employers offering equal apprenticeships for young women and men and ring-fence funds for training women in sectors in which they're underrepresented. Steve Rotheram, Labour Born in Kirkby, former bricklayer Steve Rotheram was a city councillor in Liverpool and also Lord Mayor during the city's European Capital of Culture year in 2008. He was also elected MP for Liverpool Walton in 2010, and re-elected to the seat in 2015. Mr Rotheram is pledging to cut the cost of the fast tag for motorists driving through the Mersey tunnels. He wants to improve education and offer better careers advice for young people, and also wants to make brownfield sites more attractive to developers. Paula Walters, UKIP Wallasey-born Paula Walters is chairman of UKIP in Wirral and lives in New Brighton with her family. She has campaigned to scrap tunnel tolls for several years. She says her local UKIP branch is one of the most thriving in the North West. A civil servant, she studied English and biomolecular science at degree-level. She has also lived in South Africa where she attended the University of Pretoria. She believes Liverpool city centre has attracted money at the expense of outlying areas, one of the things she wants to tackle.	Those hoping to become the first mayor of the Liverpool City Region have less than a month remaining in which to secure your vote.

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

compression-pairs

• Dataset: compression-pairs at 605bc91
• Size: 4,000 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 10 tokens mean: 31.89 tokens max: 125 tokens	min: 5 tokens mean: 10.21 tokens max: 28 tokens

• Samples:

  sentence1	sentence2
  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: AdaptiveLayerLoss with these parameters:

  {
      "loss": "MultipleNegativesSymmetricRankingLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

sciq_pairs

• Dataset: sciq_pairs at 2c94ad3
• Size: 6,500 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 7 tokens mean: 17.26 tokens max: 60 tokens	min: 2 tokens mean: 84.37 tokens max: 512 tokens

• Samples:

  sentence1	sentence2
  What type of organism is commonly used in preparation of foods such as cheese and yogurt?	Mesophiles grow best in moderate temperature, typically between 25°C and 40°C (77°F and 104°F). Mesophiles are often found living in or on the bodies of humans or other animals. The optimal growth temperature of many pathogenic mesophiles is 37°C (98°F), the normal human body temperature. Mesophilic organisms have important uses in food preparation, including cheese, yogurt, beer and wine.
  What phenomenon makes global winds blow northeast to southwest or the reverse in the northern hemisphere and northwest to southeast or the reverse in the southern hemisphere?	Without Coriolis Effect the global winds would blow north to south or south to north. But Coriolis makes them blow northeast to southwest or the reverse in the Northern Hemisphere. The winds blow northwest to southeast or the reverse in the southern hemisphere.
  Changes from a less-ordered state to a more-ordered state (such as a liquid to a solid) are always what?	Summary Changes of state are examples of phase changes, or phase transitions. All phase changes are accompanied by changes in the energy of a system. Changes from a more-ordered state to a less-ordered state (such as a liquid to a gas) areendothermic. Changes from a less-ordered state to a more-ordered state (such as a liquid to a solid) are always exothermic. The conversion of a solid to a liquid is called fusion (or melting). The energy required to melt 1 mol of a substance is its enthalpy of fusion (ΔHfus). The energy change required to vaporize 1 mol of a substance is the enthalpy of vaporization (ΔHvap). The direct conversion of a solid to a gas is sublimation. The amount of energy needed to sublime 1 mol of a substance is its enthalpy of sublimation (ΔHsub) and is the sum of the enthalpies of fusion and vaporization. Plots of the temperature of a substance versus heat added or versus heating time at a constant rate of heating are calledheating curves. Heating curves relate temperature changes to phase transitions. A superheated liquid, a liquid at a temperature and pressure at which it should be a gas, is not stable. A cooling curve is not exactly the reverse of the heating curve because many liquids do not freeze at the expected temperature. Instead, they form a supercooled liquid, a metastable liquid phase that exists below the normal melting point. Supercooled liquids usually crystallize on standing, or adding a seed crystal of the same or another substance can induce crystallization.

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

qasc_pairs

• Dataset: qasc_pairs at a34ba20
• Size: 6,500 training samples
• Columns: id, sentence1, and sentence2
• Approximate statistics based on the first 1000 samples:

  	id	sentence1	sentence2
  type	string	string	string
  details	min: 17 tokens mean: 21.35 tokens max: 27 tokens	min: 5 tokens mean: 11.47 tokens max: 25 tokens	min: 14 tokens mean: 35.55 tokens max: 66 tokens

• Samples:

  id	sentence1	sentence2
  3E7TUJ2EGCLQNOV1WEAJ2NN9ROPD9K	What type of water formation is formed by clouds?	beads of water are formed by water vapor condensing. Clouds are made of water vapor.. Beads of water can be formed by clouds.
  3LS2AMNW5FPNJK3C3PZLZCPX562OQO	Where do beads of water come from?	beads of water are formed by water vapor condensing. Condensation is the change of water vapor to a liquid.. Vapor turning into a liquid leaves behind beads of water
  3TMFV4NEP8DPIPCI8H9VUFHJG8V8W3	What forms beads of water?	beads of water are formed by water vapor condensing. An example of water vapor is steam.. Steam forms beads of water.

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

openbookqa_pairs

• Dataset: openbookqa_pairs at 388097e
• Size: 2,740 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 3 tokens mean: 13.83 tokens max: 78 tokens	min: 4 tokens mean: 11.37 tokens max: 30 tokens

• Samples:

  sentence1	sentence2
  The sun is responsible for	the sun is the source of energy for physical cycles on Earth
  When food is reduced in the stomach	digestion is when stomach acid breaks down food
  Stars are	a star is made of gases

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

msmarco_pairs

• Dataset: msmarco_pairs at 28ff31e
• Size: 6,500 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 4 tokens mean: 8.61 tokens max: 27 tokens	min: 18 tokens mean: 75.09 tokens max: 206 tokens

• Samples:

  sentence1	sentence2
  what are the liberal arts?	liberal arts. 1. the academic course of instruction at a college intended to provide general knowledge and comprising the arts, humanities, natural sciences, and social sciences, as opposed to professional or technical subjects.
  what is the mechanism of action of fibrinolytic or thrombolytic drugs?	Baillière's Clinical Haematology. 6 Mechanism of action of the thrombolytic agents. 6 Mechanism of action of the thrombolytic agents JEFFREY I. WEITZ Fibrin formed during the haemostatic, inflammatory or tissue repair process serves a temporary role, and must be degraded to restore normal tissue function and structure.
  what is normal plat count	78 Followers. A. Platelets are the tiny blood cells that help stop bleeding by binding together to form a clump or plug at sites of injury inside blood vessels. A normal platelet count is between 150,000 and 450,000 platelets per microliter (one-millionth of a liter, abbreviated mcL).The average platelet count is 237,000 per mcL in men and 266,000 per mcL in women.8 Followers. A. Platelets are the tiny blood cells that help stop bleeding by binding together to form a clump or plug at sites of injury inside blood vessels. A normal platelet count is between 150,000 and 450,000 platelets per microliter (one-millionth of a liter, abbreviated mcL).

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

nq_pairs

• Dataset: nq_pairs at f9e894e
• Size: 6,500 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 10 tokens mean: 11.77 tokens max: 21 tokens	min: 16 tokens mean: 131.57 tokens max: 512 tokens

• Samples:

  sentence1	sentence2
  when did richmond last play in a preliminary final	Richmond Football Club Richmond began 2017 with 5 straight wins, a feat it had not achieved since 1995. A series of close losses hampered the Tigers throughout the middle of the season, including a 5-point loss to the Western Bulldogs, 2-point loss to Fremantle, and a 3-point loss to the Giants. Richmond ended the season strongly with convincing victories over Fremantle and St Kilda in the final two rounds, elevating the club to 3rd on the ladder. Richmond's first final of the season against the Cats at the MCG attracted a record qualifying final crowd of 95,028; the Tigers won by 51 points. Having advanced to the first preliminary finals for the first time since 2001, Richmond defeated Greater Western Sydney by 36 points in front of a crowd of 94,258 to progress to the Grand Final against Adelaide, their first Grand Final appearance since 1982. The attendance was 100,021, the largest crowd to a grand final since 1986. The Crows led at quarter time and led by as many as 13, but the Tigers took over the game as it progressed and scored seven straight goals at one point. They eventually would win by 48 points – 16.12 (108) to Adelaide's 8.12 (60) – to end their 37-year flag drought.[22] Dustin Martin also became the first player to win a Premiership medal, the Brownlow Medal and the Norm Smith Medal in the same season, while Damien Hardwick was named AFL Coaches Association Coach of the Year. Richmond's jump from 13th to premiers also marked the biggest jump from one AFL season to the next.
  who sang what in the world's come over you	Jack Scott (singer) At the beginning of 1960, Scott again changed record labels, this time to Top Rank Records.[1] He then recorded four Billboard Hot 100 hits – "What in the World's Come Over You" (#5), "Burning Bridges" (#3) b/w "Oh Little One" (#34), and "It Only Happened Yesterday" (#38).[1] "What in the World's Come Over You" was Scott's second gold disc winner.[6] Scott continued to record and perform during the 1960s and 1970s.[1] His song "You're Just Gettin' Better" reached the country charts in 1974.[1] In May 1977, Scott recorded a Peel session for BBC Radio 1 disc jockey, John Peel.
  who produces the most wool in the world	Wool Global wool production is about 2 million tonnes per year, of which 60% goes into apparel. Wool comprises ca 3% of the global textile market, but its value is higher owing to dying and other modifications of the material.[1] Australia is a leading producer of wool which is mostly from Merino sheep but has been eclipsed by China in terms of total weight.[30] New Zealand (2016) is the third-largest producer of wool, and the largest producer of crossbred wool. Breeds such as Lincoln, Romney, Drysdale, and Elliotdale produce coarser fibers, and wool from these sheep is usually used for making carpets.

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

trivia_pairs

• Dataset: trivia_pairs at a7c36e3
• Size: 6,500 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 8 tokens mean: 15.16 tokens max: 48 tokens	min: 19 tokens mean: 456.87 tokens max: 512 tokens

• Samples:

  sentence1	sentence2
  Which American-born Sinclair won the Nobel Prize for Literature in 1930?	The Nobel Prize in Literature 1930 The Nobel Prize in Literature 1930 Sinclair Lewis The Nobel Prize in Literature 1930 Sinclair Lewis Prize share: 1/1 The Nobel Prize in Literature 1930 was awarded to Sinclair Lewis "for his vigorous and graphic art of description and his ability to create, with wit and humour, new types of characters". Photos: Copyright © The Nobel Foundation Share this: To cite this page MLA style: "The Nobel Prize in Literature 1930". Nobelprize.org. Nobel Media AB 2014. Web. 18 Jan 2017. http://www.nobelprize.org/nobel_prizes/literature/laureates/1930/
  Where in England was Dame Judi Dench born?	Judi Dench - IMDb IMDb Actress
  In which decade did Billboard magazine first publish and American hit chart?	The US Billboard song chart The US Billboard song chart Search this site with Google Song chart US Billboard The Billboard magazine has published various music charts starting (with sheet music) in 1894, the first "Music Hit Parade" was published in 1936 , the first "Music Popularity Chart" was calculated in 1940 . These charts became less irregular until the weekly "Hot 100" was started in 1958 . The current chart combines sales, airplay and downloads. A music collector that calls himself Bullfrog has been consolidating the complete chart from 1894 to the present day. he has published this information in a comprehenive spreadsheet (which can be obtained at bullfrogspond.com/ ). The Bullfrog data assigns each song a unique identifier, something like "1968_076" (which just happens to be the Bee Gees song "I've Gotta Get A Message To You"). This "Whitburn Number" is provided to match with the books of Joel Whitburn and consists of the year and a ranking within the year. A song that first entered the charts in December and has a long run is listed the following year. This numbering scheme means that songs which are still in the charts cannot be assigned a final id, because their ranking might change. So the definitive listing for a year cannot be final until about April. In our listing we only use songs with finalised IDs, this means that every year we have to wait until last year's entries are finalised before using them. (Source bullfrogspond.com/ , the original version used here was 20090808 with extra data from: the 2009 data from 20091219 the 2010 data from 20110305 the 2011 data from 20120929 the 2012 data from 20130330 the 2013 data from 20150328 The 20150328 data was the last one produced before the Billboard company forced the data to be withdrawn. As far as we know there are no more recent data sets available. This pattern of obtaining the data for a particular year in the middle of the following one comes from the way that the Bullfrog project generates the identifier for a song (what they call the "Prefix" in the spreadsheet). Recent entries are identified with keys like "2015-008" while older ones have keys like "2013_177". In the second case the underscore is significant, it indicates that this was the 177th biggest song released in 2013. Now, of course, during the year no one knows where a particular song will rank, so the underscore names can't be assigned until every song from a particular year has dropped out of the charts, so recent records are temporarily assigned a name with a dash. In about May of the following year the rankings are calculated and the final identifiers are assigned. That is why we at the Turret can only grab this data retrospectively. Attributes The original spreadsheet has a number of attributes, we have limited our attention to just a few of them: 134 9 The songs with the most entries on the chart were White Christmas (with 33 versions and a total of 110 weeks) and Stardust (with 19 and a total of 106 weeks). position The peak position that songs reached in the charts should show an smooth curve from number one down to the lowest position. This chart has more songs in the lower peak positions than one would expect. Before 1991 the profile of peak positions was exactly as you would expect, that year Billboard introduced the concept of "Recurrent" tracks, that is they removed any track from the chart which had spent more than twenty weeks in the chart and had fallen to the lower positions. weeks The effect of the "Recurrent" process, by which tracks are removed if they have spent at least twenty weeks in the chart and have fallen to the lower reaches, can clearly be seen in the strange spike in this attribute. This "adjustment" was intended to promote newer songs and ensure the chart does not become "stale". In fact since it was introduced in 1991 the length of long chart runs has increased, this might reflect the more conscious efforts of record companies to "game" the charts by controlling release times and promotions, or it coul

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

quora_pairs

• Dataset: quora_pairs at 451a485
• Size: 4,000 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 6 tokens mean: 13.53 tokens max: 42 tokens	min: 6 tokens mean: 13.68 tokens max: 43 tokens

• Samples:

  sentence1	sentence2
  Astrology: I am a Capricorn Sun Cap moon and cap rising...what does that say about me?	I'm a triple Capricorn (Sun, Moon and ascendant in Capricorn) What does this say about me?
  How can I be a good geologist?	What should I do to be a great geologist?
  How do I read and find my YouTube comments?	How can I see all my Youtube comments?

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

gooaq_pairs

• Dataset: gooaq_pairs at b089f72
• Size: 6,500 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 8 tokens mean: 11.6 tokens max: 21 tokens	min: 13 tokens mean: 57.74 tokens max: 127 tokens

• Samples:

  sentence1	sentence2
  is toprol xl the same as metoprolol?	Metoprolol succinate is also known by the brand name Toprol XL. It is the extended-release form of metoprolol. Metoprolol succinate is approved to treat high blood pressure, chronic chest pain, and congestive heart failure.
  are you experienced cd steve hoffman?	The Are You Experienced album was apparently mastered from the original stereo UK master tapes (according to Steve Hoffman - one of the very few who has heard both the master tapes and the CDs produced over the years). ... The CD booklets were a little sparse, but at least they stayed true to the album's original design.
  how are babushka dolls made?	Matryoshka dolls are made of wood from lime, balsa, alder, aspen, and birch trees; lime is probably the most common wood type. ... After cutting, the trees are stripped of most of their bark, although a few inner rings of bark are left to bind the wood and keep it from splitting.

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

Evaluation Datasets

nli-pairs

• Dataset: nli-pairs at d482672
• Size: 750 evaluation samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 5 tokens mean: 17.61 tokens max: 51 tokens	min: 4 tokens mean: 9.71 tokens max: 29 tokens

• Samples:

  anchor	positive
  Two women are embracing while holding to go packages.	Two woman are holding packages.
  Two young children in blue jerseys, one with the number 9 and one with the number 2 are standing on wooden steps in a bathroom and washing their hands in a sink.	Two kids in numbered jerseys wash their hands.
  A man selling donuts to a customer during a world exhibition event held in the city of Angeles	A man selling donuts to a customer.

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

scitail-pairs-pos

• Dataset: scitail-pairs-pos at 0cc4353
• Size: 750 evaluation samples
• Columns: sentence1, sentence2, and label
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	label
  type	string	string	int
  details	min: 5 tokens mean: 22.43 tokens max: 61 tokens	min: 8 tokens mean: 15.3 tokens max: 36 tokens	0: ~50.00% 1: ~50.00%

• Samples:

  sentence1	sentence2	label
  An introduction to atoms and elements, compounds, atomic structure and bonding, the molecule and chemical reactions.	Replace another in a molecule happens to atoms during a substitution reaction.	0
  Wavelength The distance between two consecutive points on a sinusoidal wave that are in phase;	Wavelength is the distance between two corresponding points of adjacent waves called.	1
  humans normally have 23 pairs of chromosomes.	Humans typically have 23 pairs pairs of chromosomes.	1

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "GISTEmbedLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

qnli-contrastive

• Dataset: qnli-contrastive at bcdcba7
• Size: 750 evaluation samples
• Columns: sentence1, sentence2, and label
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	label
  type	string	string	int
  details	min: 6 tokens mean: 14.15 tokens max: 36 tokens	min: 4 tokens mean: 36.98 tokens max: 225 tokens	0: 100.00%

• Samples:

  sentence1	sentence2	label
  What came into force after the new constitution was herald?	As of that day, the new constitution heralding the Second Republic came into force.	0
  What is the first major city in the stream of the Rhine?	The most important tributaries in this area are the Ill below of Strasbourg, the Neckar in Mannheim and the Main across from Mainz.	0
  What is the minimum required if you want to teach in Canada?	In most provinces a second Bachelor's Degree such as a Bachelor of Education is required to become a qualified teacher.	0

• Loss: AdaptiveLayerLoss with these parameters:

  {
      "loss": "OnlineContrastiveLoss",
      "n_layers_per_step": -1,
      "last_layer_weight": 1.5,
      "prior_layers_weight": 0.15,
      "kl_div_weight": 2,
      "kl_temperature": 2
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 28
• per_device_eval_batch_size: 18
• learning_rate: 2e-05
• weight_decay: 5e-07
• num_train_epochs: 4
• lr_scheduler_type: cosine_with_restarts
• lr_scheduler_kwargs: {'num_cycles': 5}
• warmup_ratio: 0.25
• save_safetensors: False
• fp16: True
• push_to_hub: True
• hub_model_id: bobox/DeBERTaV3-small-SenTra-AdaptiveLayers-AllSoft-HighTemp-n
• hub_strategy: checkpoint
• 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: 28
• per_device_eval_batch_size: 18
• 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: 5e-07
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 4
• max_steps: -1
• lr_scheduler_type: cosine_with_restarts
• lr_scheduler_kwargs: {'num_cycles': 5}
• warmup_ratio: 0.25
• warmup_steps: 0
• log_level: passive
• log_level_replica: warning
• log_on_each_node: True
• logging_nan_inf_filter: True
• save_safetensors: False
• 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: 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: True
• resume_from_checkpoint: None
• hub_model_id: bobox/DeBERTaV3-small-SenTra-AdaptiveLayers-AllSoft-HighTemp-n
• hub_strategy: checkpoint
• 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: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	scitail-pairs-pos loss	qnli-contrastive loss	nli-pairs loss	sts-test_spearman_cosine
0.1003	281	8.4339	-	-	-	-
0.2006	562	6.8644	-	-	-	-
0.3009	843	5.1225	-	-	-	-
0.4001	1121	-	2.4070	4.2827	3.6032	-
0.4011	1124	3.9997	-	-	-	-
0.5014	1405	3.6186	-	-	-	-
0.6017	1686	3.259	-	-	-	-
0.7020	1967	3.1712	-	-	-	-
0.8001	2242	-	1.6090	2.5195	2.2851	-
0.8023	2248	3.104	-	-	-	-
0.9026	2529	2.8549	-	-	-	-
1.0029	2810	2.8668	-	-	-	-
1.1031	3091	2.7466	-	-	-	-
1.2002	3363	-	1.3474	2.2222	1.8491	-
1.2034	3372	2.6502	-	-	-	-
1.3037	3653	2.2191	-	-	-	-
1.4040	3934	2.2311	-	-	-	-
1.5043	4215	2.22	-	-	-	-
1.6003	4484	-	1.2671	1.7964	1.6444	-
1.6046	4496	2.1372	-	-	-	-
1.7049	4777	2.2219	-	-	-	-
1.8051	5058	2.2618	-	-	-	-
1.9054	5339	1.9995	-	-	-	-
2.0004	5605	-	1.2434	1.8182	1.5385	-
2.0057	5620	1.9757	-	-	-	-
2.1060	5901	2.0401	-	-	-	-
2.2063	6182	1.9818	-	-	-	-
2.3066	6463	1.7816	-	-	-	-
2.4004	6726	-	1.0396	1.5587	1.5077	-
2.4069	6744	1.9239	-	-	-	-
2.5071	7025	2.0148	-	-	-	-
2.6074	7306	1.9629	-	-	-	-
2.7077	7587	1.7316	-	-	-	-
2.8005	7847	-	1.0507	1.3294	1.4039	-
2.8080	7868	1.7794	-	-	-	-
2.9083	8149	1.7029	-	-	-	-
3.0086	8430	1.7996	-	-	-	-
3.1089	8711	1.9379	-	-	-	-
3.2006	8968	-	0.9949	1.3678	1.3436	-
3.2091	8992	1.844	-	-	-	-
3.3094	9273	1.358	-	-	-	-
3.4097	9554	1.5104	-	-	-	-
3.5100	9835	1.6964	-	-	-	-
3.6006	10089	-	0.9538	1.1866	1.3098	-
3.6103	10116	1.7661	-	-	-	-
3.7106	10397	1.6529	-	-	-	-
3.8108	10678	1.6835	-	-	-	-
3.9111	10959	1.35	-	-	-	-
4.0	11208	-	-	-	-	0.5551

Framework Versions

• Python: 3.10.13
• Sentence Transformers: 3.0.1
• Transformers: 4.41.2
• PyTorch: 2.1.2
• Accelerate: 0.30.1
• Datasets: 2.19.2
• 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",
}

AdaptiveLayerLoss

@misc{li20242d,
    title={2D Matryoshka Sentence Embeddings}, 
    author={Xianming Li and Zongxi Li and Jing Li and Haoran Xie and Qing Li},
    year={2024},
    eprint={2402.14776},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

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},
}

GISTEmbedLoss

@misc{solatorio2024gistembed,
    title={GISTEmbed: Guided In-sample Selection of Training Negatives for Text Embedding Fine-tuning}, 
    author={Aivin V. Solatorio},
    year={2024},
    eprint={2402.16829},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}