ksg-dfci/TrialSpace-0825

SentenceTransformer

This is a sentence-transformers model distilled from Snowflake/snowflake-arctic-embed-l-v2.0 . It maps sentences & paragraphs to a 1024-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
• Maximum Sequence Length: 8192 tokens
• Output Dimensionality: 1024 dimensions
• Similarity Function: Cosine Similarity

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': 8192, 'do_lower_case': False, 'architecture': 'XLMRobertaModel'})
  (1): Pooling({'word_embedding_dimension': 1024, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, '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): Normalize()
)

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
queries = [
    "Cancer type: Conventional high\u2011grade (grade\u202f3) chondrosarcoma of the right scapula  \nHistology: Pleomorphic sarcoma with chondroid differentiation (high\u2011grade conventional chondrosarcoma)  \nCurrent extent: Stage\u00a0IV/metastatic \u2013 pulmonary metastases (numerous bilateral nodules, stable) and progressive hepatic metastasis (now 4.2\u202fcm, causing biliary obstruction)  \n\nBiomarkers:  \n- IDH2 p.R172S mutation (initial NGS low VAF, later VAF\u202f\u2248\u202f18\u202f%)  \n- TP53 p.R282W loss\u2011of\u2011function missense mutation  \n- CDKN2A homozygous deletion (p16 loss)  \n- MDM2 amplification (\u2248\u202f12\u2011fold)  \n- FGFR2 amplification  \n- COL2A1 p.R1060C (variant of uncertain significance)  \n- Microsatellite stability (MS\u2011Stable)  \n- PD\u2011L1 IHC 0\u202f% TPS (negative)  \n- Tumor mutational burden \u2248\u202f9\u202fmut/Mb (moderate)\n\nTreatment history:  \n# 2020\u201111\u201123 onward: Diagnosis confirmed by pathology. Initiated local external beam radiotherapy to the scapular primary (exact dates not specified, commenced shortly after diagnosis).  \n# Approx. 2022\u201109\u2011xx to 2022\u201110\u2011xx: Off\u2011label regorafenib, initiated for systemic control; duration\u202f~\u202f4\u202fweeks. Discontinued because of Grade\u202f3 hypertension and Grade\u202f3 hand\u2011foot skin reaction.  \n# Early 2022 (approximately February\u2013March): Additional systemic therapy attempted (unspecified agent) but halted \u201csix weeks ago\u201d as of 2022\u201105\u201111 due to lack of radiographic response and toxicity. Exact drug unknown; recorded as therapy discontinued.  \n# Ongoing: Supportive care with opioid analgesics for scapular pain; no further active anticancer therapy after regorafenib cessation given poor performance status (ECOG\u202f\u2265\u202f3) and limited expected benefit.",
]
documents = [
    '4. Cancer type allowed: solid tumor (all tissues). Histology allowed: any malignant histologic subtype provided the tumor harbours an IDH1 mutation. Cancer burden allowed: advanced or metastatic disease refractory to conventional therapeutic options or intolerant of such therapy; radiographically measurable disease not previously subjected to radiation, chemo‑embolisation, radio‑embolisation or other local ablative technique. Prior treatment required: evidence of refractoriness/intolerance to standard-of-care modalities. Prior treatment excluded: n/a. Biomarkers required: IDH1 gene mutation determined on local diagnostic platform (centrally reassessed retrospectively). Biomarkers excluded: n/a.',
    '1. Cancer type allowed: meningioma. Histology allowed: World Health Organization grade\u202fII or grade\u202fIII meningioma. Cancer burden allowed: recurrent or progressively growing intracranial disease with at least one meas\xadurable lesion ≥10\u202fmm on magnetic resonance imaging. Prior treatment required: prior neurosurgical resection of the index meningioma **and** prior cranial radiation therapy directed at the progressing tumour. Prior treatment excluded: • more than two distinct courses of radiation therapy administered for meningioma • a documented clinical diagnosis of Neurofibroma\xadtosis type\u202f2 OR a molecularly identified NF2 alteration • three or more prior systemic chemotherapy regimens delivered for meningeal disease. Biomarkers required: none noted. Biomarkers excluded: neurofi bromas tos i s\u202ftype\u202f2 (clinical or molecular).',
    '1. Cancer type allowed: renal cell carcinoma. Histology allowed: clear cell renal cell carcinoma. Cancer burden allowed: advanced or metastatic disease. Prior treatment required: none. Prior treatment excluded: any prior systemic therapy for advanced or metastatic renal cell carcinoma; prior belzutifan or other HIF‑2α inhibitor; prior cabozantinib.  ',
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# [1, 1024] [3, 1024]

# Get the similarity scores for the embeddings
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[0.8427, 0.1946, 0.3175]])

Training Details

Training Datasets

Unnamed Dataset

• Size: 415,029 training samples
• Columns: sentence_0, sentence_1, and label
• Approximate statistics based on the first 1000 samples:

  	sentence_0	sentence_1	label
  type	string	string	float
  details	min: 200 tokens mean: 569.76 tokens max: 1231 tokens	min: 14 tokens mean: 121.13 tokens max: 459 tokens	min: 0.0 mean: 0.73 max: 1.0

• Samples:

  sentence_0	sentence_1	label
  Cancer type: Epithelioid sarcoma (classic “distal” type) Histology: High‑grade (grade 2) epithelioid sarcoma with loss of INI1 (SMARCB1) expression Current extent: Metastatic disease (bilateral pulmonary nodules and a solitary hepatic lesion) Biomarkers: - SMARCB1 biallelic deletion → complete loss of nuclear INI1 (predictive for EZH2 inhibition) - PTEN truncating mutation (loss of function) - KRAS p.G13D activating mutation (confers MAPK pathway activation, predicts resistance to EGFR‐directed agents) - CDKN2A homozygous deletion - TP53 splice‑site variant (non‑functional p53) - Low‑level EGFR copy number gain (≈ 3 copies) Treatment history: # 2017‑mid – early 2018: Neoadjuvant doxorubicin × 4 cycles (anthracycline chemotherapy) – administered prior to definitive local therapy (exact dates not specified) # Early 2018 (approx. Mar‑Apr 2018): External beam radiation therapy, 45 Gy in 25 fractions to the left hand (definitive EBRT) – completed by 2018‑04‑21 (acu...	3. Cancer type allowed: tumours deficient in INI1 (or otherwise INI1‑negative/aberrant) irrespective of organ origin, and any solid tumour harbouring an activating (“gain‑of‑function”) mutation in EZH2. Histology allowed: diverse solid‐tumour histologies meeting the INI1‑loss definition or carrying an EZH2 GOF point mutation/amplification. Cancer burden allowed: metastatic disease or unresectable locally advanced disease that is relapsed or refractory after prior therapy (including cases progressing within six months before enrolment). Prior treatment required: any preceding anticancer therapy whose residual toxicities have resolved to ≤grade 1. Prior treatment excluded: none stated specifically for these cohorts. Biomarkers required: (a) loss of INI1 protein by IHC or bi‑biallelic INI1 loss/mutation confirmed genetically, or (b) demonstrable EZH2 gain‑of‑function mutation/amplification identified by validated molecular platform. Biomarkers excluded: none.	1.0
  Cancer type: Non‑small cell lung cancer (lung adenocarcinoma) Histology: Moderately differentiated invasive adenocarcinoma, TTF‑1 +, Napsin‑A + Current extent: Metastatic (thoracic disease stable, solitary treated cerebellar brain metastasis, no other extracranial sites identified) Biomarkers: • EGFR exon 19 sensiti­zing deletion (present on all specimens) • PD‑L1 Tumor Proportion Score ≈30 % (initial PET/CT report) • Acquired EGFR C797S substitution (NGS 2022‑08‑20) • MET copy number amplification ≈9 copies (clinical note 2023‑05‑18) • Tumor mutational burden 4.2 Mut/Mb (low) • Microsatellite stable / proficient mismatch repair (NGS 2022‑08‑20) Treatment history: # 2020‑05‑12 to 2020‑07‑30: Neoadjuvant cisplatin 75 mg/m² + pemetrexed 500 mg/m² q21 days – three cycles completed (last dose approx June 2021). Best response: partial reduction of primary tumor (≈44 % decrease on 2021‑04‑12 imaging). # Late 2021 (post‑neoadjuvant): VATS left lower‑lobe wedge resect...	1. Cancer type allowed: non small cell lung cancer. Histology allowed: non small cell lung cancer. Cancer burden allowed: advanced unresectable or metastatic disease. Prior treatment required: ≤ five prior anticancer regimens permissible. Prior treatment excluded: prior anti‑programmed death receptor 1, anti‑programmed death ligand 1, or anti‑programmed death ligand 2 antibody exposure. Biomarkers required: ☐. Biomarkers excluded: ☐.	1.0
  Cancer type: Urinary bladder urothelial carcinoma Histology: High‑grade papillary urothelial carcinoma, WHO/ISUT Grade III Current extent: Metastatic (progressive disease with FDG‑avid left supraclavicular lymph node and suspected pulmonary involvement; persistent pelvic nodal disease) Biomarkers: • FGFG3 S249C mutation • TP53 R248W mutation • CDKN2A loss • ERBB2 amplification (copy number = 6) • PIK3CA E545K mutation • KDM6A truncating alteration • MDM2 amplification • STAG2 truncating alteration • TERT promoter −124 C>T mutation • Low PD‑L1 tumor proportion score ≈ 5% Treatment history: # 2012‑11‑21 → Initial transurethral resection of bladder tumor (TURBT) showing high‑grade papillary urothelial carcinoma invading muscularis propria (pT2). # Early 2013 → Neoadjuvant MVAC chemotherapy (≥3 cycles reported by 09‑15‑2013; total 4 cycles completed by October 2013). Partial radiographic response of primary lesion, stable pelvic nodal disease. # 10‑13‑20...	11. Cancer type allowed: Urothelial/bladder cancer. Histology allowed: transitional cell carcinoma. Cancer burden allowed: advanced/metastatic disease. Prior treatment required: none specific. Prior treatment excluded: none beyond the ubiquitous recent‑therapy ban. Biomarkers required: none. Biomarkers excluded: none.	1.0

• Loss: OnlineContrastiveLoss

Unnamed Dataset

• Size: 309,687 training samples
• Columns: sentence_0 and sentence_1
• Approximate statistics based on the first 1000 samples:

  	sentence_0	sentence_1
  type	string	string
  details	min: 196 tokens mean: 568.43 tokens max: 1095 tokens	min: 15 tokens mean: 124.96 tokens max: 452 tokens

• Samples:

  sentence_0	sentence_1
  Cancer type: Colorectal adenocarcinoma Histology: Moderately differentiated invasive adenocarcinoma, Grade 2 Current extent: Metastatic (initially stage IV with hepatic metastases and a solitary right frontal brain metastasis; brain lesion treated with stereotactic radiosurgery and shows complete radiographic remission; hepatic disease persists with intermittent progression) Biomarkers: KRAS wild‑type; Microsatellite stable (MS‑stable); Tumor mutational burden low (≈4 mutations/Mb); IDH2 R172K activating mutation (VAF ~8%); additional somatic alterations – APC truncation, TP53 missense, SMAD4 loss, MYC/CDK6 amplifications; no germline‑relevant BRCA1/2 changes Treatment history: # 2013‑05‑20: Right hemicolectomy (curative resection of primary colonic tumor) # 2013‑05‑20: Stereotactic radiosurgery to solitary right frontal brain metastasis (single fraction 18 Gy) – resulted in complete radiographic remission, no residual neurological symptoms # 2013‑01‑03 to 2013‑05‑20: F...	1. Cancer type allowed: various solid malignant neoplasms excluding central nervous system. Histology allowed: diverse solid tumor histologies. Cancer burden allowed: advanced or metastatic disease, recurrent or progressive after standard therapy. Prior treatment required: disease must have progressed after receipt of standard therapeutic regimen (which may include targeted therapy against mutant IDH1/IDH2). Prior treatment excluded: none specified beyond what is covered under exclusion criteria. Biomarkers required: presence of an IDH1 and/or IDH2 gene mutation (to be determined using archival tumor specimen or fresh biopsy). Biomarkers excluded: none.
  Cancer type: Breast cancer Histology: Invasive (ductal) carcinoma, NST, grade 2, hormone‑receptor‑positive (ER⁺/PR⁺), HER2‑negative, right breast Current extent: Metastatic (initially bone‐only, later hepatic progression; disease remains active/metastatic as of latest note in 02/2014) Biomarkers: - Estrogen receptor positive, Progesterone receptor positive, HER2 negative (by IHC/FISH at diagnosis) - ESR1 Y537S mutation (detected 09/2013) – associated with resistance to aromatase inhibitors - PIK3CA E545K activating mutation (detected 09/2013) - Microsatellite instability stable (MSI‑S) (09/2013) - Tumor mutational burden 8 mutations/Mb (intermediate) (09/2013) Treatment history: # 02/03/2012 – ~05/07/2012: Letrozole 2.5 mg orally once daily (first‑line aromatase inhibitor). Achieved stable bone disease on serial scans but met RECIST criteria for radiographic progression (new hepatic lesion) after ≈4 months. # ~05/2013 – early 02/2014: Everolimus combined with L...	2. Cancer type allowed: metastatic breast cancer. Histology allowed: invasive breast adenocarcinoma. Cancer burden allowed: hormonereceptor‑positive, human epidermal growth factor receptor 2‑negative metastatic disease; absence of currently active or symptomatic central nervous system involvement. Prior treatment required: endocrine‑resistance demonstrated; receipt of at least one line containing combined hormonal therapy together with an FDA‑approved cyclin‑dependent kinase 4/6 inhibitor; total number of prior systemic regimens for locoregionally unresectable/metastatic disease limited to ≥ 1 and ≤ 4. Prior treatment excluded: n/a. Biomarkers required: estrogen‑receptor and/or progesterone‑receptor positivity; HER2 negativity verified according to guideline testing methodology. Biomarkers excluded: n/a.
  Cancer type: Ewing sarcoma Histology: Small round blue cell tumor, CD99⁺, NKX2.2⁺, EWSR1‑FLI1 fusion-positive Current extent: Metastatic (persistent FDG‑avid L2 vertebral body lesion, stable disease) Biomarkers: - Confirmed EWSR1‑FLI1 translocation (detected 2017‑06‑01 pathology, reconfirmed 2020‑05‑08 NGS) - CDK4 amplification (high level, 2020‑05‑08 NGS) - CCND1 copy‑number gain (modest, 2020‑05‑08 NGS) - TP53 p.R175H missense mutation (NGS Jan 2025) - ATRX splice‑variant alteration (NGS Jan 2025) - STAG2 frameshift loss‑of‑function (NGS Jan 2025) Treatment history: # 2017‑06‑01 to 2018‑04‑xx: Neoadjuvant interval‑compressed VDC/IE (vincristine, doxorubicin, cyclophosphamide alternating with ifosfamide/etoposide) – 8 cycles total (partial response observed on 2018‑05‑23 MRI, ~40% shrinkage) # 2018‑09‑29 to 2018‑09‑29: Left scapular partial scapulectomy with prosthetic reconstruction (negative surgical margins) # 2018‑10‑09 to 2018‑10‑09: Adjuvant external‑...	4. Cancer type allowed: Ewing sarcoma. Histology allowed: classic Ewing sarcoma (small round blue cell tumour) with characteristic marker profile (CD99 positive, keratin variable, INI1 retained) confirming diagnosis. Cancer burden allowed: metastatic disease or unresectable locally advanced disease that is relapsed or refractory after prior therapy. Prior treatment required: any preceding anticantic therapy whose residual toxicities have resolved to ≤grade 1. Prior treatment excluded: none stated specifically for this cohort. Biomarkers required: morphological features compatible with Ewing sarcoma together with supportive immunoprofile; no additional molecular prerequisite stipulated for entry into this exploratory cohort. Biomarkers excluded: none.

• Loss: MultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim",
      "gather_across_devices": false
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• per_device_train_batch_size: 6
• per_device_eval_batch_size: 6
• num_train_epochs: 2
• multi_dataset_batch_sampler: round_robin

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: no
• prediction_loss_only: True
• per_device_train_batch_size: 6
• per_device_eval_batch_size: 6
• 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
• num_train_epochs: 2
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.0
• 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
• hub_revision: None
• 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
• 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
• liger_kernel_config: None
• eval_use_gather_object: False
• average_tokens_across_devices: False
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: round_robin
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Click to expand

Epoch	Step	Training Loss
0.0194	500	0.4392
0.0387	1000	0.4558
0.0581	1500	0.457
0.0775	2000	0.4577
0.0969	2500	0.4412
0.1162	3000	0.4488
0.1356	3500	0.4486
0.1550	4000	0.4616
0.1744	4500	0.4728
0.1937	5000	0.4619
0.2131	5500	0.4545
0.2325	6000	0.4495
0.2519	6500	0.4618
0.2712	7000	0.4214
0.2906	7500	0.4412
0.3100	8000	0.4505
0.3294	8500	0.4313
0.3487	9000	0.4508
0.3681	9500	0.4303
0.3875	10000	0.4399
0.4069	10500	0.448
0.4262	11000	0.4408
0.4456	11500	0.4337
0.4650	12000	0.4273
0.4843	12500	0.4385
0.5037	13000	0.4437
0.5231	13500	0.439
0.5425	14000	0.4284
0.5618	14500	0.4214
0.5812	15000	0.4238
0.6006	15500	0.4225
0.6200	16000	0.4187
0.6393	16500	0.4151
0.6587	17000	0.4383
0.6781	17500	0.4243
0.6975	18000	0.4148
0.7168	18500	0.419
0.7362	19000	0.4169
0.7556	19500	0.4184
0.7750	20000	0.4191
0.7943	20500	0.4329
0.8137	21000	0.4339
0.8331	21500	0.4087
0.8524	22000	0.4161
0.8718	22500	0.4242
0.8912	23000	0.4183
0.9106	23500	0.4076
0.9299	24000	0.4095
0.9493	24500	0.4328
0.9687	25000	0.4114
0.9881	25500	0.4242
1.0074	26000	0.4158
1.0268	26500	0.3909
1.0462	27000	0.3999
1.0656	27500	0.4025
1.0849	28000	0.4115
1.1043	28500	0.3843
1.1237	29000	0.4177
1.1431	29500	0.4083
1.1624	30000	0.4025
1.1818	30500	0.4133
1.2012	31000	0.4006
1.2206	31500	0.3985
1.2399	32000	0.3999
1.2593	32500	0.394
1.2787	33000	0.3927
1.2980	33500	0.3964
1.3174	34000	0.4001
1.3368	34500	0.3956
1.3562	35000	0.3899
1.3755	35500	0.388
1.3949	36000	0.3867
1.4143	36500	0.3982
1.4337	37000	0.394
1.4530	37500	0.3942
1.4724	38000	0.3913
1.4918	38500	0.3909
1.5112	39000	0.3757
1.5305	39500	0.3829
1.5499	40000	0.3874
1.5693	40500	0.3883
1.5887	41000	0.3783
1.6080	41500	0.4041
1.6274	42000	0.403
1.6468	42500	0.3806
1.6662	43000	0.3825
1.6855	43500	0.3944
1.7049	44000	0.3956
1.7243	44500	0.382
1.7436	45000	0.3911
1.7630	45500	0.3823
1.7824	46000	0.3771
1.8018	46500	0.3784
1.8211	47000	0.3853
1.8405	47500	0.3864
1.8599	48000	0.3724
1.8793	48500	0.3856
1.8986	49000	0.3862
1.9180	49500	0.376
1.9374	50000	0.377
1.9568	50500	0.3937
1.9761	51000	0.3819
1.9955	51500	0.3899

Framework Versions

• Python: 3.12.3
• Sentence Transformers: 5.1.0
• Transformers: 4.55.4
• PyTorch: 2.7.1+cu126
• Accelerate: 1.10.0
• Datasets: 4.0.0
• Tokenizers: 0.21.4

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",
}

MultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}