mertcobanov/mpnet-base-all-nli-triplet-turkish-v3

MPNet base trained on AllNLI-turkish triplets

This is a sentence-transformers model finetuned from microsoft/mpnet-base on the all-nli-triplets-turkish dataset. It maps sentences & paragraphs to a 768-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: microsoft/mpnet-base
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • all-nli-triplets-turkish
• Language: en
• License: apache-2.0

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: MPNetModel 
  (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("mertcobanov/mpnet-base-all-nli-triplet-turkish-v3")
# Run inference
sentences = [
    'Ağaçlarla çevrili bulvar denize üç bloktan daha az uzanıyor.',
    'Deniz üç sokak bile uzakta değil.',
    'Denize ulaşmak için caddeden iki mil yol almanız gerekiyor.',
]
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

Triplet

• Datasets: all-nli-dev-turkish and all-nli-test-turkish
• Evaluated with TripletEvaluator

Metric	all-nli-dev-turkish	all-nli-test-turkish
cosine_accuracy	0.7423	0.7503

Training Details

Training Dataset

all-nli-triplets-turkish

• Dataset: all-nli-triplets-turkish at bff203b
• Size: 13,842 training samples
• Columns: anchor_translated, positive_translated, and negative_translated
• Approximate statistics based on the first 1000 samples:

  	anchor_translated	positive_translated	negative_translated
  type	string	string	string
  details	min: 8 tokens mean: 13.42 tokens max: 95 tokens	min: 8 tokens mean: 31.64 tokens max: 93 tokens	min: 6 tokens mean: 32.03 tokens max: 89 tokens

• Samples:

  anchor_translated	positive_translated	negative_translated
  Asyalı okul çocukları birbirlerinin omuzlarında oturuyor.	Okul çocukları bir arada	Asyalı fabrika işçileri oturuyor.
  İnsanlar dışarıda.	Arka planda resmi kıyafetler giymiş bir grup insan var ve beyaz gömlekli, haki pantolonlu bir adam toprak yoldan yeşil çimenlere atlıyor.	Bir odada üç kişiyle birlikte büyük bir kamera tutan bir adam.
  Bir adam dışarıda.	Adam yarış sırasında yan sepetten bir su birikintisine düşer.	Beyaz bir sarık sarmış gömleksiz bir adam bir ağaç gövdesine tırmanıyor.

• Loss: MultipleNegativesRankingLoss with these parameters:

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

Evaluation Dataset

all-nli-triplets-turkish

• Dataset: all-nli-triplets-turkish at bff203b
• Size: 6,584 evaluation samples
• Columns: anchor_translated, positive_translated, and negative_translated
• Approximate statistics based on the first 1000 samples:

  	anchor_translated	positive_translated	negative_translated
  type	string	string	string
  details	min: 5 tokens mean: 42.62 tokens max: 192 tokens	min: 5 tokens mean: 22.58 tokens max: 77 tokens	min: 5 tokens mean: 22.07 tokens max: 65 tokens

• Samples:

  anchor_translated	positive_translated	negative_translated
  Ayrıca, bu özel tüketim vergileri, diğer vergiler gibi, hükümetin ödeme zorunluluğunu sağlama yetkisini kullanarak belirlenir.	Hükümetin ödeme zorlaması, özel tüketim vergilerinin nasıl hesaplandığını belirler.	Özel tüketim vergileri genel kuralın bir istisnasıdır ve aslında GSYİH payına dayalı olarak belirlenir.
  Gri bir sweatshirt giymiş bir sanatçı, canlı renklerde bir kasaba tablosu üzerinde çalışıyor.	Bir ressam gri giysiler içinde bir kasabanın resmini yapıyor.	Bir kişi bir beyzbol sopası tutuyor ve gelen bir atış için planda bekliyor.
  İmkansız.	Yapılamaz.	Tamamen mümkün.

• Loss: MultipleNegativesRankingLoss with these parameters:

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

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• learning_rate: 2e-05
• num_train_epochs: 10
• warmup_ratio: 0.1
• fp16: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 16
• per_device_eval_batch_size: 16
• 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: 2e-05
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 10
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.1
• warmup_steps: 0
• log_level: passive
• log_level_replica: warning
• log_on_each_node: True
• logging_nan_inf_filter: True
• save_safetensors: True
• save_on_each_node: False
• save_only_model: False
• restore_callback_states_from_checkpoint: False
• no_cuda: False
• use_cpu: False
• use_mps_device: False
• seed: 42
• data_seed: None
• jit_mode_eval: False
• use_ipex: False
• bf16: False
• fp16: True
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• local_rank: 0
• ddp_backend: None
• tpu_num_cores: None
• tpu_metrics_debug: False
• debug: []
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_prefetch_factor: None
• past_index: -1
• disable_tqdm: False
• remove_unused_columns: True
• label_names: None
• load_best_model_at_end: 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: False
• hub_always_push: False
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• include_for_metrics: []
• eval_do_concat_batches: True
• fp16_backend: auto
• push_to_hub_model_id: None
• push_to_hub_organization: None
• mp_parameters:
• auto_find_batch_size: False
• full_determinism: False
• torchdynamo: None
• ray_scope: last
• ddp_timeout: 1800
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• dispatch_batches: None
• split_batches: None
• include_tokens_per_second: False
• include_num_input_tokens_seen: False
• neftune_noise_alpha: None
• optim_target_modules: None
• batch_eval_metrics: False
• eval_on_start: False
• use_liger_kernel: False
• eval_use_gather_object: False
• average_tokens_across_devices: False
• prompts: None
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	Validation Loss	all-nli-dev-turkish_cosine_accuracy	all-nli-test-turkish_cosine_accuracy
0	0	-	-	0.6092	-
0.1155	100	3.3654	2.9084	0.6624	-
0.2309	200	2.6321	1.7277	0.7395	-
0.3464	300	1.9629	1.5000	0.7512	-
0.4619	400	1.6662	1.4965	0.7494	-
0.5774	500	1.4712	1.5374	0.7418	-
0.6928	600	1.0429	1.6301	0.7360	-
0.8083	700	0.8995	2.1626	0.7044	-
0.9238	800	0.7269	2.0440	0.6996	-
1.0381	900	1.0584	1.6714	0.7438	-
1.1536	1000	1.1864	1.5326	0.7495	-
1.2691	1100	1.0193	1.4498	0.7518	-
1.3845	1200	0.8237	1.5399	0.7506	-
1.5	1300	0.8279	1.6747	0.7521	-
1.6155	1400	0.626	1.5776	0.7453	-
1.7309	1500	0.5396	1.8877	0.7139	-
1.8464	1600	0.4294	2.2258	0.6947	-
1.9619	1700	0.4988	1.8753	0.7204	-
2.0762	1800	0.6987	1.5408	0.7524	-
2.1917	1900	0.6684	1.4434	0.7618	-
2.3072	2000	0.6072	1.4840	0.7520	-
2.4226	2100	0.5081	1.5225	0.7561	-
2.5381	2200	0.5216	1.5280	0.7514	-
2.6536	2300	0.2627	1.8830	0.7227	-
2.7691	2400	0.2585	1.9529	0.7221	-
2.8845	2500	0.129	2.2323	0.7047	-
3.0	2600	0.1698	2.2904	0.7063	-
3.1143	2700	0.5559	1.6110	0.7553	-
3.2298	2800	0.4356	1.5544	0.7508	-
3.3453	2900	0.3886	1.5437	0.7539	-
3.4607	3000	0.3573	1.6262	0.7539	-
3.5762	3100	0.2652	1.8391	0.7321	-
3.6917	3200	0.0765	2.0359	0.7186	-
3.8072	3300	0.0871	2.0946	0.7262	-
3.9226	3400	0.0586	2.2168	0.7093	-
4.0370	3500	0.1755	1.7567	0.7462	-
4.1524	3600	0.3397	1.7735	0.7442	-
4.2679	3700	0.3067	1.7475	0.7497	-
4.3834	3800	0.246	1.7075	0.7476	-
4.4988	3900	0.253	1.7648	0.7483	-
4.6143	4000	0.1223	1.9139	0.7246	-
4.7298	4100	0.0453	2.1138	0.7152	-
4.8453	4200	0.0241	2.2354	0.7240	-
4.9607	4300	0.0363	2.3080	0.7251	-
5.0751	4400	0.1897	1.7394	0.7494	-
5.1905	4500	0.2114	1.6929	0.7524	-
5.3060	4600	0.2101	1.7402	0.7556	-
5.4215	4700	0.1471	1.7990	0.7445	-
5.5370	4800	0.1783	1.8060	0.7456	-
5.6524	4900	0.0215	2.0118	0.7325	-
5.7679	5000	0.0083	2.0766	0.7265	-
5.8834	5100	0.0138	2.2054	0.7201	-
5.9988	5200	0.0144	2.1667	0.7164	-
6.1132	5300	0.2023	1.7309	0.7543	-
6.2286	5400	0.1356	1.6685	0.7622	-
6.3441	5500	0.1307	1.7292	0.7527	-
6.4596	5600	0.1222	1.8403	0.7435	-
6.5751	5700	0.1049	1.8456	0.7394	-
6.6905	5800	0.0051	1.9898	0.7362	-
6.8060	5900	0.0131	2.0532	0.7310	-
6.9215	6000	0.0132	2.2237	0.7186	-
7.0358	6100	0.0453	1.8965	0.7397	-
7.1513	6200	0.1109	1.7195	0.7550	-
7.2667	6300	0.1002	1.7547	0.7530	-
7.3822	6400	0.0768	1.7701	0.7433	-
7.4977	6500	0.0907	1.8472	0.7406	-
7.6132	6600	0.038	1.9162	0.7377	-
7.7286	6700	0.0151	1.9407	0.7312	-
7.8441	6800	0.0087	1.9657	0.7289	-
7.9596	6900	0.0104	2.0302	0.7227	-
8.0739	7000	0.0727	1.8692	0.7514	-
8.1894	7100	0.0733	1.8039	0.7520	-
8.3048	7200	0.0728	1.7400	0.7539	-
8.4203	7300	0.0537	1.8062	0.7461	-
8.5358	7400	0.059	1.8469	0.7489	-
8.6513	7500	0.0089	1.9033	0.7403	-
8.7667	7600	0.0034	1.9683	0.7354	-
8.8822	7700	0.0018	2.0075	0.7366	-
8.9977	7800	0.0023	2.0646	0.7322	-
9.1120	7900	0.0642	1.9063	0.7430	-
9.2275	8000	0.0596	1.8492	0.7468	-
9.3430	8100	0.0479	1.8180	0.7517	-
9.4584	8200	0.0561	1.8122	0.7468	-
9.5739	8300	0.0311	1.8528	0.7456	-
9.6894	8400	0.0069	1.8778	0.7447	-
9.8048	8500	0.0027	1.8989	0.7423	-
9.9203	8600	0.0093	1.9089	0.7423	-
9.9896	8660	-	-	-	0.7503

Framework Versions

• Python: 3.10.14
• Sentence Transformers: 3.3.1
• Transformers: 4.46.3
• PyTorch: 2.3.0
• Accelerate: 1.1.1
• Datasets: 3.1.0
• Tokenizers: 0.20.3

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

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}
}