Bheri/ithasa-mmbert-v3

SentenceTransformer based on jhu-clsp/mmBERT-base

This is a sentence-transformers model finetuned from jhu-clsp/mmBERT-base. 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: jhu-clsp/mmBERT-base
• Maximum Sequence Length: 128 tokens
• Output Dimensionality: 768 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': 128, 'do_lower_case': False, 'architecture': 'ModernBertModel'})
  (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("sentence_transformers_model_id")
# Run inference
sentences = [
    'attenuated vaccines:',
    'कम संवेदनशील टीकेः',
    '६.५% दसादशे',
]
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, 0.3723, 0.1543],
#         [0.3723, 1.0000, 0.2746],
#         [0.1543, 0.2746, 1.0000]])

Evaluation

Metrics

Translation

• Dataset: eval-en-sa
• Evaluated with TranslationEvaluator

Metric	Value
src2trg_accuracy	0.616
trg2src_accuracy	0.604
mean_accuracy	0.61

Training Details

Training Dataset

Unnamed Dataset

• Size: 3,749,530 training samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 12 tokens mean: 31.26 tokens max: 88 tokens	min: 19 tokens mean: 67.93 tokens max: 128 tokens

• Samples:

  sentence1	sentence2
  There was no Mughal tradition of primogeniture, the systematic passing of rule, upon an emperor's death, to his eldest son.	चक्रवर्तिनः मृत्योः अनन्तरं तस्य शासनस्य व्यवस्थितरूपेण सङ्क्रमणस्य, मुघलपरम्परायाः ज्येष्ठपुत्राधिकारपद्धतिः नासीत्।
  The four sons of Shah Jahan all held governorships during their father's reign.	शाह्-जहाँ-नामकस्य चत्वारः पुत्राः, सर्वे पितुः शासनकाले शासकपदम् अधारयन्।
  In this regard he discusses the correlation between social opportunities of education and health and how both of these complement economic and political freedoms as a healthy and well-educated person is better suited to make informed economic decisions and be involved in fruitful political demonstrations etc.	अस्मिन् विषये सः शिक्षणस्य स्वास्थ्यस्य च सामाजिकावकाशानाम् अन्योन्य-सम्बन्धस्य, तथा च एतद्द्वयम् अपि आर्थिक-राजनैतिक-स्वातन्त्र्ययोः कथं पूरकं भवतः इति च चर्चां करोति, यतोहि स्वस्था सुशिक्षिता च व्यक्तिः ज्ञानपूर्वम् आर्थिकविषयान् निर्णेतुं तथा फलप्रदेषु राजनैतिकेषु प्रतिपादनादिषु संलग्नः भवितुं च अधिकारी भवति इति।

• Loss: MultipleNegativesRankingLoss with these parameters:

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

Evaluation Dataset

Unnamed Dataset

• Size: 1,000 evaluation samples
• Columns: sentence1 and sentence2
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2
  type	string	string
  details	min: 5 tokens mean: 11.9 tokens max: 67 tokens	min: 5 tokens mean: 23.13 tokens max: 128 tokens

• Samples:

  sentence1	sentence2
  plus 2 tempered glass screen protectors:	6 पश्चात तापाभिसंतप्तॊ विदुर समार कर्शितः
  "Take sadaqah (alms) from their wealth in order to purify them with it." (p.	अप्येकाङ्गेऽप्यधोवस्तुमिच्छामि च सुकुत्सिते" ॥
  "Who could it possibly be?"	कश्च तासेः सम्भवति ?

• 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: 32
• num_train_epochs: 5
• max_steps: 12000
• learning_rate: 2e-05
• warmup_steps: 500
• gradient_accumulation_steps: 4
• bf16: True
• eval_strategy: steps
• load_best_model_at_end: True

All Hyperparameters

Click to expand

• per_device_train_batch_size: 32
• num_train_epochs: 5
• max_steps: 12000
• learning_rate: 2e-05
• lr_scheduler_type: linear
• lr_scheduler_kwargs: None
• warmup_steps: 500
• optim: adamw_torch_fused
• optim_args: None
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• optim_target_modules: None
• gradient_accumulation_steps: 4
• average_tokens_across_devices: True
• max_grad_norm: 1.0
• label_smoothing_factor: 0.0
• bf16: True
• fp16: False
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• use_liger_kernel: False
• liger_kernel_config: None
• use_cache: False
• neftune_noise_alpha: None
• torch_empty_cache_steps: None
• auto_find_batch_size: False
• log_on_each_node: True
• logging_nan_inf_filter: True
• include_num_input_tokens_seen: no
• log_level: passive
• log_level_replica: warning
• disable_tqdm: False
• project: huggingface
• trackio_space_id: trackio
• eval_strategy: steps
• per_device_eval_batch_size: 8
• prediction_loss_only: True
• eval_on_start: False
• eval_do_concat_batches: True
• eval_use_gather_object: False
• eval_accumulation_steps: None
• include_for_metrics: []
• batch_eval_metrics: False
• save_only_model: False
• save_on_each_node: False
• enable_jit_checkpoint: False
• push_to_hub: False
• hub_private_repo: None
• hub_model_id: None
• hub_strategy: every_save
• hub_always_push: False
• hub_revision: None
• load_best_model_at_end: True
• ignore_data_skip: False
• restore_callback_states_from_checkpoint: False
• full_determinism: False
• seed: 42
• data_seed: None
• use_cpu: False
• 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
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• dataloader_prefetch_factor: None
• remove_unused_columns: True
• label_names: None
• train_sampling_strategy: random
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• ddp_backend: None
• ddp_timeout: 1800
• fsdp: []
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• deepspeed: None
• debug: []
• skip_memory_metrics: True
• do_predict: False
• resume_from_checkpoint: None
• warmup_ratio: None
• local_rank: -1
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Click to expand

Epoch	Step	Training Loss	Validation Loss	eval-en-sa_mean_accuracy
0.0034	100	3.1353	-	-
0.0068	200	2.7273	-	-
0.0102	300	1.8263	-	-
0.0137	400	1.1810	-	-
0.0171	500	0.8952	-	-
0.0205	600	0.7068	-	-
0.0239	700	0.5979	-	-
0.0273	800	0.5412	-	-
0.0307	900	0.5255	-	-
0.0341	1000	0.4847	0.2013	0.5045
0.0376	1100	0.4752	-	-
0.0410	1200	0.4645	-	-
0.0444	1300	0.4173	-	-
0.0478	1400	0.4220	-	-
0.0512	1500	0.4163	-	-
0.0546	1600	0.3978	-	-
0.0580	1700	0.3895	-	-
0.0614	1800	0.3778	-	-
0.0649	1900	0.3904	-	-
0.0683	2000	0.3656	0.1436	0.563
0.0717	2100	0.3565	-	-
0.0751	2200	0.3526	-	-
0.0785	2300	0.3632	-	-
0.0819	2400	0.3468	-	-
0.0853	2500	0.3506	-	-
0.0888	2600	0.3505	-	-
0.0922	2700	0.3466	-	-
0.0956	2800	0.3422	-	-
0.0990	2900	0.3393	-	-
0.1024	3000	0.3345	0.1240	0.587
0.1058	3100	0.3238	-	-
0.1092	3200	0.3230	-	-
0.1127	3300	0.3281	-	-
0.1161	3400	0.3246	-	-
0.1195	3500	0.3111	-	-
0.1229	3600	0.3092	-	-
0.1263	3700	0.3187	-	-
0.1297	3800	0.3293	-	-
0.1331	3900	0.3246	-	-
0.1366	4000	0.3174	0.1165	0.598
0.1400	4100	0.3213	-	-
0.1434	4200	0.3167	-	-
0.1468	4300	0.3142	-	-
0.1502	4400	0.3070	-	-
0.1536	4500	0.3094	-	-
0.1570	4600	0.3084	-	-
0.1604	4700	0.3068	-	-
0.1639	4800	0.3060	-	-
0.1673	4900	0.3020	-	-
0.1707	5000	0.3072	0.1133	0.6045
0.1741	5100	0.3151	-	-
0.1775	5200	0.3121	-	-
0.1809	5300	0.3059	-	-
0.1843	5400	0.3069	-	-
0.1878	5500	0.3069	-	-
0.1912	5600	0.3134	-	-
0.1946	5700	0.3017	-	-
0.1980	5800	0.3088	-	-
0.2014	5900	0.3011	-	-
0.2048	6000	0.3075	0.1109	0.608
0.2082	6100	0.2957	-	-
0.2117	6200	0.3049	-	-
0.2151	6300	0.2994	-	-
0.2185	6400	0.2951	-	-
0.2219	6500	0.3116	-	-
0.2253	6600	0.3155	-	-
0.2287	6700	0.2938	-	-
0.2321	6800	0.2824	-	-
0.2355	6900	0.2973	-	-
0.2390	7000	0.3111	0.1100	0.6065
0.2424	7100	0.2973	-	-
0.2458	7200	0.2995	-	-
0.2492	7300	0.2962	-	-
0.2526	7400	0.2994	-	-
0.2560	7500	0.2964	-	-
0.2594	7600	0.2997	-	-
0.2629	7700	0.2932	-	-
0.2663	7800	0.2993	-	-
0.2697	7900	0.2987	-	-
0.2731	8000	0.2898	0.1084	0.6085
0.2765	8100	0.3007	-	-
0.2799	8200	0.2935	-	-
0.2833	8300	0.2885	-	-
0.2868	8400	0.3021	-	-
0.2902	8500	0.2958	-	-
0.2936	8600	0.3056	-	-
0.2970	8700	0.2908	-	-
0.3004	8800	0.3096	-	-
0.3038	8900	0.2924	-	-
0.3072	9000	0.3019	0.1077	0.607
0.3107	9100	0.2985	-	-
0.3141	9200	0.2906	-	-
0.3175	9300	0.2961	-	-
0.3209	9400	0.3044	-	-
0.3243	9500	0.3005	-	-
0.3277	9600	0.2943	-	-
0.3311	9700	0.2948	-	-
0.3345	9800	0.3046	-	-
0.3380	9900	0.2948	-	-
0.3414	10000	0.3060	0.1083	0.608
0.3448	10100	0.2906	-	-
0.3482	10200	0.2958	-	-
0.3516	10300	0.2919	-	-
0.3550	10400	0.3041	-	-
0.3584	10500	0.3055	-	-
0.3619	10600	0.2975	-	-
0.3653	10700	0.2984	-	-
0.3687	10800	0.2883	-	-
0.3721	10900	0.2949	-	-
0.3755	11000	0.2987	0.1083	0.6085
0.3789	11100	0.2938	-	-
0.3823	11200	0.2942	-	-
0.3858	11300	0.2879	-	-
0.3892	11400	0.2909	-	-
0.3926	11500	0.2899	-	-
0.3960	11600	0.2921	-	-
0.3994	11700	0.2944	-	-
0.4028	11800	0.2985	-	-
0.4062	11900	0.3027	-	-
0.4097	12000	0.2988	0.1082	0.61

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.10.18
• Sentence Transformers: 5.2.3
• Transformers: 5.2.0
• PyTorch: 2.8.0+cu128
• Accelerate: 1.12.0
• Datasets: 3.3.2
• Tokenizers: 0.22.1

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",
}

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}
}