This is a sentence-transformers model finetuned from dbmdz/distilbert-base-turkish-cased. 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: dbmdz/distilbert-base-turkish-cased
Maximum Sequence Length: 512 tokens
Output Dimensionality: 768 dimensions
Similarity Function: Cosine Similarity

Language: tr
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: 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})
)

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.

python

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("orhanxakarsu/sentence-distilbert-turkish")
# Run inference
sentences = [
    "İki kadın, Çin'deki bir markette bir ürüne bakıyor.",
    'Alışveriş yapan iki kadın',
    'Kadınlar bir spor salonunda çalışıyorlar.',
]
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

Dataset: all-nli-turkish-dev
Evaluated with TripletEvaluator

Metric	Value
cosine_accuracy	0.9802

Training Details

Training Dataset

Unnamed Dataset

Size: 814,596 training samples
Columns: anchor, positive, and negative
Approximate statistics based on the first 1000 samples:
anchor positive negative
type string string string
details
min: 3 tokens
mean: 18.16 tokens
max: 91 tokens
min: 4 tokens
mean: 10.54 tokens
max: 136 tokens
min: 4 tokens
mean: 10.73 tokens
max: 29 tokens

Samples:

anchor	positive	negative
`Beyaz gömlekli ve güneş gözlüklü bir kadın, kucağında bir bebekle dışarıda bir sandalyede oturuyor.`	`Bebek yerden yukarıda oturuyor`	`Adam bir top atıyor`
`Mavi yakalı gömlek giyen ve kazaklı bir adam ve beyaz gömlek giyen hasır şapka takan bir kadın.`	`Yan yana bir erkek ve bir kadın var.`	`Evli bir çift akşam yemeği yiyor.`
`Adam içeride.`	`Siyah fötr şapkalı bir adam bir arenada boğaya biniyor.`	`Yeşil üniforma giyen beş subayla birlikte taş bir binanın önünde cep telefonuyla konuşan bir papaz; ikisi ayakta, diğerleri oturuyor.`

Loss: MultipleNegativesRankingLoss with these parameters:
json
1{ 2 "scale": 20.0, 3 "similarity_fct": "cos_sim" 4}

Evaluation Dataset

Unnamed Dataset

Size: 8,229 evaluation samples
Columns: anchor, positive, and negative
Approximate statistics based on the first 1000 samples:
anchor positive negative
type string string string
details
min: 4 tokens
mean: 17.91 tokens
max: 80 tokens
min: 4 tokens
mean: 10.62 tokens
max: 35 tokens
min: 4 tokens
mean: 11.01 tokens
max: 33 tokens

Samples:

anchor	positive	negative
`Patlamanın büyüklüğünün güçlü bir örneği, Haragosha Tapınağı'nda bulunur, burada tapınağın kemerinin üst crosebar'ını görebilirsiniz, geri kalanı sertleşmiş lav tarafından batırılmıştır.`	`Patlamanın büyüklüğünün sonucu Haragosha Tapınağı'nda görülüyor.`	`Haragosha Tapınağı bu güne kadar tamamen sağlamdır.`
`Arkeolojik kazı yapan iki kişi.`	`Kazı yapan insanlar var.`	`Kimse kazmıyor.`
`İşçiler, Martins'in ünlü Louisiana sosis satıcısı çadırının önünde sıraya giren müşterilere hizmet veriyor`	`Müşteriler bir satıcı çadırının önünde sıraya giriyor.`	`Pamuk şeker yiyen bir grup insan var.`

Loss: MultipleNegativesRankingLoss with these parameters:
json
1{ 2 "scale": 20.0, 3 "similarity_fct": "cos_sim" 4}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: steps
per_device_train_batch_size: 64
per_device_eval_batch_size: 64
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: 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: 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
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
eval_use_gather_object: False
prompts: None
batch_sampler: no_duplicates
multi_dataset_batch_sampler: proportional

Training Logs

Click to expand

Epoch	Step	Training Loss	Validation Loss	all-nli-turkish-dev_cosine_accuracy
0	0	-	-	0.5808
0.0786	1000	3.5327	1.9481	0.7607
0.1571	2000	1.5833	1.2787	0.8260
0.2357	3000	1.2338	1.0960	0.8533
0.3142	4000	1.1031	0.9897	0.8695
0.3928	5000	0.998	0.9077	0.8793
0.4714	6000	0.9412	0.8434	0.8914
0.5499	7000	0.8703	0.7904	0.8982
0.6285	8000	0.8094	0.7311	0.9068
0.7070	9000	0.7653	0.6894	0.9086
0.7856	10000	0.7248	0.6509	0.9162
0.8642	11000	0.673	0.6145	0.9205
0.9427	12000	0.6514	0.5762	0.9273
1.0213	13000	0.6259	0.5463	0.9334
1.0999	14000	0.5874	0.5276	0.9332
1.1784	15000	0.5518	0.5053	0.9366
1.2570	16000	0.5277	0.4783	0.9391
1.3355	17000	0.5075	0.4571	0.9419
1.4141	18000	0.4906	0.4379	0.9454
1.4927	19000	0.475	0.4234	0.9465
1.5712	20000	0.447	0.4046	0.9499
1.6498	21000	0.4307	0.3908	0.9508
1.7283	22000	0.4126	0.3773	0.9548
1.8069	23000	0.3985	0.3654	0.9564
1.8855	24000	0.3748	0.3582	0.9560
1.9640	25000	0.3675	0.3449	0.9581
2.0426	26000	0.3545	0.3390	0.9586
2.1211	27000	0.3456	0.3335	0.9595
2.1997	28000	0.3295	0.3255	0.9626
2.2783	29000	0.3198	0.3146	0.9624
2.3568	30000	0.3107	0.3101	0.9642
2.4354	31000	0.3139	0.3014	0.9665
2.5139	32000	0.2982	0.3005	0.9659
2.5925	33000	0.2903	0.2891	0.9663
2.6711	34000	0.2778	0.2859	0.9662
2.7496	35000	0.2731	0.2812	0.9667
2.8282	36000	0.2613	0.2757	0.9677
2.9067	37000	0.2566	0.2680	0.9689
2.9853	38000	0.2488	0.2674	0.9699
3.0639	39000	0.2434	0.2594	0.9694
3.1424	40000	0.2375	0.2574	0.9705
3.2210	41000	0.2295	0.2553	0.9706
3.2996	42000	0.223	0.2501	0.9703
3.3781	43000	0.2209	0.2455	0.9719
3.4567	44000	0.2211	0.2409	0.9711
3.5352	45000	0.2097	0.2396	0.9728
3.6138	46000	0.2068	0.2345	0.9734
3.6924	47000	0.1994	0.2298	0.9731
3.7709	48000	0.1986	0.2299	0.9730
3.8495	49000	0.1878	0.2271	0.9728
3.9280	50000	0.1872	0.2244	0.9739
4.0066	51000	0.1821	0.2249	0.9734
4.0852	52000	0.1823	0.2188	0.9739
4.1637	53000	0.1736	0.2176	0.9748
4.2423	54000	0.1691	0.2152	0.9745
4.3208	55000	0.1665	0.2148	0.9753
4.3994	56000	0.1663	0.2133	0.9748
4.4780	57000	0.1666	0.2123	0.9755
4.5565	58000	0.1589	0.2082	0.9758
4.6351	59000	0.155	0.2053	0.9762
4.7136	60000	0.155	0.2037	0.9762
4.7922	61000	0.1536	0.2031	0.9764
4.8708	62000	0.1443	0.2020	0.9759
4.9493	63000	0.146	0.1999	0.9752
5.0279	64000	0.1417	0.1969	0.9764
5.1064	65000	0.1407	0.1966	0.9761
5.1850	66000	0.1342	0.1981	0.9757
5.2636	67000	0.1342	0.1933	0.9768
5.3421	68000	0.1312	0.1944	0.9758
5.4207	69000	0.1329	0.1932	0.9772
5.4993	70000	0.1304	0.1908	0.9768
5.5778	71000	0.1247	0.1880	0.9772
5.6564	72000	0.1221	0.1861	0.9779
5.7349	73000	0.1225	0.1831	0.9784
5.8135	74000	0.1205	0.1854	0.9790
5.8921	75000	0.1152	0.1815	0.9789
5.9706	76000	0.1161	0.1827	0.9782
6.0492	77000	0.1151	0.1819	0.9781
6.1277	78000	0.113	0.1818	0.9780
6.2063	79000	0.1102	0.1823	0.9784
6.2849	80000	0.1067	0.1798	0.9780
6.3634	81000	0.1067	0.1782	0.9790
6.4420	82000	0.1116	0.1779	0.9782
6.5205	83000	0.107	0.1752	0.9782
6.5991	84000	0.1039	0.1739	0.9792
6.6777	85000	0.1013	0.1728	0.9789
6.7562	86000	0.1029	0.1713	0.9786
6.8348	87000	0.0972	0.1721	0.9791
6.9133	88000	0.0991	0.1703	0.9790
6.9919	89000	0.0955	0.1708	0.9791
7.0705	90000	0.097	0.1715	0.9786
7.1490	91000	0.0941	0.1716	0.9793
7.2276	92000	0.0922	0.1712	0.9795
7.3062	93000	0.0921	0.1706	0.9789
7.3847	94000	0.091	0.1691	0.9793
7.4633	95000	0.0942	0.1689	0.9787
7.5418	96000	0.0905	0.1678	0.9790
7.6204	97000	0.0871	0.1664	0.9792
7.6990	98000	0.0859	0.1666	0.9793
7.7775	99000	0.0876	0.1656	0.9785
7.8561	100000	0.084	0.1643	0.9795
7.9346	101000	0.0853	0.1654	0.9795
8.0132	102000	0.083	0.1640	0.9789
8.0918	103000	0.0849	0.1637	0.9795
8.1703	104000	0.0816	0.1626	0.9797
8.2489	105000	0.0803	0.1627	0.9796
8.3274	106000	0.0802	0.1623	0.9796
8.4060	107000	0.0808	0.1622	0.9798
8.4846	108000	0.0836	0.1632	0.9792
8.5631	109000	0.0791	0.1612	0.9796
8.6417	110000	0.0761	0.1609	0.9798
8.7202	111000	0.0782	0.1604	0.9797
8.7988	112000	0.0784	0.1604	0.9803
8.8774	113000	0.0737	0.1600	0.9804
8.9559	114000	0.0762	0.1602	0.9799
9.0345	115000	0.0764	0.1597	0.9802
9.1130	116000	0.0761	0.1600	0.9799
9.1916	117000	0.0729	0.1592	0.9797
9.2702	118000	0.0728	0.1595	0.9803
9.3487	119000	0.0722	0.1590	0.9798
9.4273	120000	0.0745	0.1591	0.9797
9.5059	121000	0.0741	0.1591	0.9798
9.5844	122000	0.0715	0.1587	0.9797
9.6630	123000	0.0719	0.1581	0.9799
9.7415	124000	0.0716	0.1578	0.9799
9.8201	125000	0.0714	0.1582	0.9801
9.8987	126000	0.0712	0.1579	0.9803
9.9772	127000	0.0707	0.1581	0.9802

Framework Versions

Python: 3.12.4
Sentence Transformers: 3.3.1
Transformers: 4.44.2
PyTorch: 2.4.1+cu124
Accelerate: 0.33.0
Datasets: 3.1.0
Tokenizers: 0.19.1

Citation

BibTeX

Sentence Transformers

bibtex

@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

bibtex

@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}
}

	anchor	positive	negative
type	string	string	string
details	min: 3 tokens mean: 18.16 tokens max: 91 tokens	min: 4 tokens mean: 10.54 tokens max: 136 tokens	min: 4 tokens mean: 10.73 tokens max: 29 tokens