gte-multilingual-base-arabic-triplets

This is a sentence-transformers model finetuned from Alibaba-NLP/gte-multilingual-base on the silma-arabic-triplets-dataset-v1.0 dataset. It maps sentences & paragraphs to a 768-dimensional dense vector space and can be used for retrieval.

Model Details

Model Description

Model Type: Sentence Transformer
Base model: Alibaba-NLP/gte-multilingual-base
Maximum Sequence Length: 8192 tokens
Output Dimensionality: 768 dimensions
Similarity Function: Cosine Similarity
Supported Modality: Text
Training Dataset:
- silma-arabic-triplets-dataset-v1.0
Language: ar
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({'transformer_task': 'feature-extraction', 'modality_config': {'text': {'method': 'forward', 'method_output_name': 'last_hidden_state'}}, 'module_output_name': 'token_embeddings', 'architecture': 'NewModel'})
  (1): Pooling({'embedding_dimension': 768, 'pooling_mode': 'cls', '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("RamzyBakir/arabic-gte-multilingual-embed-medium")
# Run inference
sentences = [
    'هم نوع من مثل القطط.',
    'إنهم يشبهون القطط',
    'إنسان مع قطة',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.8787, 0.6596],
#         [0.8787, 1.0000, 0.7666],
#         [0.6596, 0.7666, 1.0000]])

Evaluation

Metrics

Triplet

Datasets: arabic-triplet-fast and arabic-triplet-eval-full
Evaluated with TripletEvaluator

Metric	arabic-triplet-fast	arabic-triplet-eval-full
cosine_accuracy	0.9804	0.9824

Training Details

Training Dataset

silma-arabic-triplets-dataset-v1.0

Dataset: silma-arabic-triplets-dataset-v1.0
Size: 2,027,899 training samples
Columns: anchor, positive, and negative

Approximate statistics based on the first 1000 samples:

	anchor	positive	negative
type	string	string	string
details	min: 5 tokens mean: 19.32 tokens max: 80 tokens	min: 5 tokens mean: 16.68 tokens max: 104 tokens	min: 5 tokens mean: 16.28 tokens max: 77 tokens

Samples:

anchor	positive	negative
`They think the web is a force for good, and most don’t want governments to regulate it.`	`وهم يعتقدون أن شبكة الويب هي قوة تستخدم للخير، ولا يرغب أغلبهم في أن تقوم الحكومات بتنظيم تلك الشبكة.`	`أصبحت الحكومة أكثر حساسية لتأثير الإنترنت على السياسة الداخلية وسنت قوانين تزيد من سلطتها لتنظيم هذا القطاع.`
`طفل صغير يرتدي نظارات زرقاء يجلس على طوف في بركة سباحة.`	`طفل صغير يجلس على طوف في حمام سباحة ويرتدي نظارات زرقاء.`	`امرأة ترقص بعنف على خشبة المسرح`
`امرأة وطفل يسيرون على الرصيف المغطى بالأوراق متجهين نحو شخصين يركبان خيول.`	`سيدة وطفل يسيرون على الرصيف باتجاه الخيول`	`امرأة وطفلان يسيران في حديقة.`

Loss: CachedMultipleNegativesRankingLoss with these parameters:

{
    "scale": 50.0,
    "similarity_fct": "cos_sim",
    "mini_batch_size": 64,
    "gather_across_devices": false,
    "directions": [
        "query_to_doc"
    ],
    "partition_mode": "joint",
    "hardness_mode": null,
    "hardness_strength": 0.0
}

Evaluation Dataset

silma-arabic-triplets-dataset-v1.0

Dataset: silma-arabic-triplets-dataset-v1.0
Size: 10,000 evaluation samples
Columns: anchor, positive, and negative

Approximate statistics based on the first 1000 samples:

	anchor	positive	negative
type	string	string	string
details	min: 5 tokens mean: 19.09 tokens max: 91 tokens	min: 5 tokens mean: 16.36 tokens max: 90 tokens	min: 5 tokens mean: 16.43 tokens max: 84 tokens

Samples:

anchor	positive	negative
`It is typically called the Constitution of the Fifth Republic, and replaced that of the Fourth Republic dating from 1946.`	`ويسمى عادة دستور الجمهورية الخامسة، وحل محل دستور الجمهورية الرابع الذي يعود تاريخه إلى عام 1946.`	`في نوفمبر 1946، اعتمدت الجمعية الوطنية أول دستور للجمهورية.`
`ما هو أفضل يوم في حياتك؟`	`أي يوم في حياتك اعتبرته أفضل يوم في حياتك؟`	`ما هي وصفات الملفات؟`
`She married French researcher Jean Ghata.`	`تزوجت من الباحث الفرنسي جان غاتا.`	`ثم انتقلوا إلى فرنسا حيث تزوجوا.`

Loss: CachedMultipleNegativesRankingLoss with these parameters:

{
    "scale": 50.0,
    "similarity_fct": "cos_sim",
    "mini_batch_size": 64,
    "gather_across_devices": false,
    "directions": [
        "query_to_doc"
    ],
    "partition_mode": "joint",
    "hardness_mode": null,
    "hardness_strength": 0.0
}

Training Hyperparameters

Non-Default Hyperparameters

per_device_train_batch_size: 128
per_device_eval_batch_size: 256
learning_rate: 2e-05
max_steps: 15842
lr_scheduler_type: cosine_with_min_lr
lr_scheduler_kwargs: {'min_lr': 1e-06}
warmup_ratio: 0.05
bf16: True
dataloader_num_workers: 4
remove_unused_columns: False
load_best_model_at_end: True
batch_sampler: no_duplicates

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
prediction_loss_only: True
per_device_train_batch_size: 128
per_device_eval_batch_size: 256
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: 3.0
max_steps: 15842
lr_scheduler_type: cosine_with_min_lr
lr_scheduler_kwargs: {'min_lr': 1e-06}
warmup_ratio: 0.05
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: True
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: 4
dataloader_prefetch_factor: None
past_index: -1
disable_tqdm: False
remove_unused_columns: False
label_names: None
load_best_model_at_end: True
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}
parallelism_config: None
deepspeed: None
label_smoothing_factor: 0.0
optim: adamw_torch_fused
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: no_duplicates
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}

Training Logs

Click to expand

Epoch	Step	Training Loss	Validation Loss	arabic-triplet-fast_cosine_accuracy	arabic-triplet-eval-full_cosine_accuracy
0.0001	1	0.5997	-	-	-
0.0063	100	0.4092	-	-	-
0.0126	200	0.3724	-	-	-
0.0189	300	0.3321	-	-	-
0.0252	400	0.3078	-	-	-
0.0316	500	0.2871	-	-	-
0.0379	600	0.2782	-	-	-
0.0442	700	0.2669	-	-	-
0.0505	800	0.2679	-	-	-
0.0568	900	0.2628	-	-	-
0.0631	1000	0.252	-	-	-
0.0694	1100	0.2424	-	-	-
0.0757	1200	0.2474	-	-	-
0.0821	1300	0.2321	-	-	-
0.0884	1400	0.2484	-	-	-
0.0947	1500	0.2309	-	-	-
0.1000	1584	-	0.3115	0.9704	-
0.1010	1600	0.2291	-	-	-
0.1073	1700	0.2242	-	-	-
0.1136	1800	0.2142	-	-	-
0.1199	1900	0.209	-	-	-
0.1262	2000	0.2259	-	-	-
0.1326	2100	0.2169	-	-	-
0.1389	2200	0.2076	-	-	-
0.1452	2300	0.2129	-	-	-
0.1515	2400	0.2041	-	-	-
0.1578	2500	0.204	-	-	-
0.1641	2600	0.2173	-	-	-
0.1704	2700	0.2086	-	-	-
0.1767	2800	0.2032	-	-	-
0.1830	2900	0.2101	-	-	-
0.1894	3000	0.2026	-	-	-
0.1957	3100	0.1977	-	-	-
0.2000	3168	-	0.2913	0.9713	-
0.2020	3200	0.1975	-	-	-
0.2083	3300	0.2019	-	-	-
0.2146	3400	0.1999	-	-	-
0.2209	3500	0.191	-	-	-
0.2272	3600	0.1996	-	-	-
0.2335	3700	0.201	-	-	-
0.2399	3800	0.1976	-	-	-
0.2462	3900	0.1963	-	-	-
0.2525	4000	0.1903	-	-	-
0.2588	4100	0.1826	-	-	-
0.2651	4200	0.1879	-	-	-
0.2714	4300	0.1764	-	-	-
0.2777	4400	0.1864	-	-	-
0.2840	4500	0.1909	-	-	-
0.2903	4600	0.1803	-	-	-
0.2967	4700	0.1819	-	-	-
0.2999	4752	-	0.2619	0.9762	-
0.3030	4800	0.187	-	-	-
0.3093	4900	0.1904	-	-	-
0.3156	5000	0.1899	-	-	-
0.3219	5100	0.1764	-	-	-
0.3282	5200	0.1828	-	-	-
0.3345	5300	0.1725	-	-	-
0.3408	5400	0.1674	-	-	-
0.3472	5500	0.1757	-	-	-
0.3535	5600	0.166	-	-	-
0.3598	5700	0.178	-	-	-
0.3661	5800	0.1765	-	-	-
0.3724	5900	0.1677	-	-	-
0.3787	6000	0.1653	-	-	-
0.3850	6100	0.176	-	-	-
0.3913	6200	0.1533	-	-	-
0.3977	6300	0.1622	-	-	-
0.3999	6336	-	0.2459	0.9771	-
0.4040	6400	0.1741	-	-	-
0.4103	6500	0.1624	-	-	-
0.4166	6600	0.1639	-	-	-
0.4229	6700	0.1674	-	-	-
0.4292	6800	0.1665	-	-	-
0.4355	6900	0.1679	-	-	-
0.4418	7000	0.1611	-	-	-
0.4481	7100	0.1661	-	-	-
0.4545	7200	0.1684	-	-	-
0.4608	7300	0.1674	-	-	-
0.4671	7400	0.1746	-	-	-
0.4734	7500	0.1684	-	-	-
0.4797	7600	0.1667	-	-	-
0.4860	7700	0.1605	-	-	-
0.4923	7800	0.1537	-	-	-
0.4986	7900	0.171	-	-	-
0.4999	7920	-	0.2387	0.9767	-
0.5050	8000	0.1587	-	-	-
0.5113	8100	0.1623	-	-	-
0.5176	8200	0.1704	-	-	-
0.5239	8300	0.1575	-	-	-
0.5302	8400	0.1671	-	-	-
0.5365	8500	0.1608	-	-	-
0.5428	8600	0.1537	-	-	-
0.5491	8700	0.1568	-	-	-
0.5555	8800	0.1582	-	-	-
0.5618	8900	0.1598	-	-	-
0.5681	9000	0.1613	-	-	-
0.5744	9100	0.1628	-	-	-
0.5807	9200	0.1507	-	-	-
0.5870	9300	0.148	-	-	-
0.5933	9400	0.1573	-	-	-
0.5996	9500	0.147	-	-	-
0.5999	9504	-	0.2270	0.9788	-
0.6059	9600	0.1502	-	-	-
0.6123	9700	0.1445	-	-	-
0.6186	9800	0.1534	-	-	-
0.6249	9900	0.1544	-	-	-
0.6312	10000	0.1509	-	-	-
0.6375	10100	0.1599	-	-	-
0.6438	10200	0.1579	-	-	-
0.6501	10300	0.1525	-	-	-
0.6564	10400	0.1371	-	-	-
0.6628	10500	0.1456	-	-	-
0.6691	10600	0.148	-	-	-
0.6754	10700	0.1472	-	-	-
0.6817	10800	0.1448	-	-	-
0.6880	10900	0.1488	-	-	-
0.6943	11000	0.1589	-	-	-
0.6999	11088	-	0.2218	0.9799	-
0.7006	11100	0.1464	-	-	-
0.7069	11200	0.1391	-	-	-
0.7132	11300	0.1489	-	-	-
0.7196	11400	0.1492	-	-	-
0.7259	11500	0.1561	-	-	-
0.7322	11600	0.1498	-	-	-
0.7385	11700	0.1553	-	-	-
0.7448	11800	0.1485	-	-	-
0.7511	11900	0.1432	-	-	-
0.7574	12000	0.1385	-	-	-
0.7637	12100	0.1497	-	-	-
0.7701	12200	0.145	-	-	-
0.7764	12300	0.1354	-	-	-
0.7827	12400	0.1345	-	-	-
0.7890	12500	0.1472	-	-	-
0.7953	12600	0.141	-	-	-
0.7998	12672	-	0.2167	0.9802	-
0.8016	12700	0.1376	-	-	-
0.8079	12800	0.1332	-	-	-
0.8142	12900	0.1469	-	-	-
0.8206	13000	0.142	-	-	-
0.8269	13100	0.1391	-	-	-
0.8332	13200	0.1512	-	-	-
0.8395	13300	0.1467	-	-	-
0.8458	13400	0.1485	-	-	-
0.8521	13500	0.1485	-	-	-
0.8584	13600	0.1412	-	-	-
0.8647	13700	0.1482	-	-	-
0.8710	13800	0.1532	-	-	-
0.8774	13900	0.1402	-	-	-
0.8837	14000	0.136	-	-	-
0.8900	14100	0.1416	-	-	-
0.8963	14200	0.1427	-	-	-
0.8998	14256	-	0.2136	0.9800	-
0.9026	14300	0.1496	-	-	-
0.9089	14400	0.1415	-	-	-
0.9152	14500	0.1395	-	-	-
0.9215	14600	0.1367	-	-	-
0.9279	14700	0.1424	-	-	-
0.9342	14800	0.1421	-	-	-
0.9405	14900	0.1312	-	-	-
0.9468	15000	0.1427	-	-	-
0.9531	15100	0.1421	-	-	-
0.9594	15200	0.1347	-	-	-
0.9657	15300	0.141	-	-	-
0.9720	15400	0.144	-	-	-
0.9784	15500	0.1417	-	-	-
0.9847	15600	0.1416	-	-	-
0.9910	15700	0.1356	-	-	-
0.9973	15800	0.1403	-	-	-
0.9998	15840	-	0.2124	0.9804	-
-1	-1	-	-	-	0.9824

The bold row denotes the saved checkpoint.

Training Time

Training: 1.1 hours

Framework Versions

Python: 3.12.6
Sentence Transformers: 5.4.1
Transformers: 4.56.0
PyTorch: 2.8.0+cu129
Accelerate: 1.10.1
Datasets: 4.8.5
Tokenizers: 0.22.0

Citation

BibTeX

Sentence Transformers

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

CachedMultipleNegativesRankingLoss

@misc{gao2021scaling,
    title={Scaling Deep Contrastive Learning Batch Size under Memory Limited Setup},
    author={Luyu Gao and Yunyi Zhang and Jiawei Han and Jamie Callan},
    year={2021},
    eprint={2101.06983},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}