LequeuISIR/ModernBERT-base-DPR-8e-05

SentenceTransformer based on answerdotai/ModernBERT-base

This is a sentence-transformers model finetuned from answerdotai/ModernBERT-base on the json 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: answerdotai/ModernBERT-base
• Maximum Sequence Length: 8192 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • json

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}) with Transformer model: 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("LequeuISIR/ModernBERT-base-DPR-8e-05")
# Run inference
sentences = [
    'This incites social hatred, threatens economic and social stability, and undermines trust in the authorities.',
    '\xa0The conditions for a healthy entrepreneurship, where the most innovative and creative win and where the source of enrichment cannot be property speculation or guilds and networks.   ',
    'As a result, the profits of the oligarchs are more than 400 times what our entire country gets from the exploitation of natural resources.',
]
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]

Training Details

Training Dataset

json

• Dataset: json
• Size: 478,146 training samples
• Columns: sentence1, sentence2, and label
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	label
  type	string	string	int
  details	min: 17 tokens mean: 33.73 tokens max: 107 tokens	min: 16 tokens mean: 33.84 tokens max: 101 tokens	0: ~57.50% 1: ~4.10% 2: ~38.40%

• Samples:

  sentence1	sentence2	label
  There have also been other important structural changes in the countryside, which have come together to form this new, as yet unknown, country.	Meanwhile, investment, which is the way to increase production, employment capacity and competitiveness of the economy, fell from 20% of output in 1974 to only 11.8% on average between 1984 and 1988.	0
  Introduce new visa categories so we can be responsive to humanitarian needs and incentivise greater investment in our domestic infrastructure and regional economies	The purpose of the project is to design and implement public policies aimed at achieving greater and faster inclusion of immigrants.	2
  and economic crimes that seriously and generally affect the fundamental rights of individuals and the international community as a whole.	For the first time in the history, not only of Ecuador, but of the entire world, a government promoted a public audit process of the foreign debt and declared some of its tranches illegitimate and immoral.	0

• Loss: CoSENTLoss with these parameters:

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

Evaluation Dataset

json

• Dataset: json
• Size: 478,146 evaluation samples
• Columns: sentence1, sentence2, and label
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	label
  type	string	string	int
  details	min: 17 tokens mean: 33.62 tokens max: 103 tokens	min: 16 tokens mean: 34.48 tokens max: 111 tokens	0: ~57.30% 1: ~2.90% 2: ~39.80%

• Samples:

  sentence1	sentence2	label
  The anchoring of the Slovak Republic in the European Union allows citizens to feel: secure politically, secure economically, secure socially.	Radikale Venstre wants Denmark to participate fully and firmly in EU cooperation on immigration, asylum and cross-border crime.	2
  Portugal's participation in the Community's negotiation of the next financial perspective should also be geared in the same direction.	Given the dynamic international framework, safeguarding the national interest requires adjustments to each of these vectors.	2
  On asylum, the Green Party will: Dismantle the direct provision system and replace it with an efficient and humane system for determining the status of asylum seekers	The crisis in the coal sector subsequently forced these immigrant workers to move into other economic sectors such as metallurgy, chemicals, construction and transport.	2

• Loss: CoSENTLoss with these parameters:

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

Training Hyperparameters

Non-Default Hyperparameters

• per_device_train_batch_size: 64
• per_device_eval_batch_size: 64
• learning_rate: 8e-05
• num_train_epochs: 5
• warmup_ratio: 0.05
• bf16: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: no
• 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: 8e-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: 5
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• 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: 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
• 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
0.0837	500	6.425
0.1673	1000	6.0308
0.2510	1500	5.9522
0.3346	2000	5.7818
0.4183	2500	5.7122
0.5019	3000	5.6378
0.5856	3500	5.5503
0.6692	4000	5.4429
0.7529	4500	5.4246
0.8365	5000	5.3536
0.9202	5500	5.4072
1.0038	6000	5.3033
1.0875	6500	4.7611
1.1712	7000	4.7535
1.2548	7500	4.7503
1.3385	8000	4.7453
1.4221	8500	4.7413
1.5058	9000	4.6753
1.5894	9500	4.67
1.6731	10000	4.7352
1.7567	10500	4.7164
1.8404	11000	4.6784
1.9240	11500	4.651
2.0077	12000	4.5708
2.0914	12500	3.6274
2.1750	13000	3.5683
2.2587	13500	3.7028
2.3423	14000	3.5859
2.4260	14500	3.6872
2.5096	15000	3.5148
2.5933	15500	3.7241
2.6769	16000	3.5983
2.7606	16500	3.6269
2.8442	17000	3.6078
2.9279	17500	3.6292
3.0115	18000	3.5151
3.0952	18500	2.5933
3.1789	19000	2.599
3.2625	19500	2.5598
3.3462	20000	2.5577
3.4298	20500	2.5827
3.5135	21000	2.5598
3.5971	21500	2.4173
3.6808	22000	2.5884
3.7644	22500	2.4313
3.8481	23000	2.5669
3.9317	23500	2.5162
4.0154	24000	2.2531
4.0990	24500	1.3758
4.1827	25000	1.5491
4.2664	25500	1.4933
4.3500	26000	1.5139
4.4337	26500	1.4607
4.5173	27000	1.6117
4.6010	27500	1.5395
4.6846	28000	1.493
4.7683	28500	1.3984
4.8519	29000	1.4183
4.9356	29500	1.3517

Framework Versions

• Python: 3.9.21
• Sentence Transformers: 3.4.0
• Transformers: 4.48.1
• PyTorch: 2.5.1+cu124
• Accelerate: 1.3.0
• Datasets: 3.2.0
• Tokenizers: 0.21.0

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",
}

CoSENTLoss

@online{kexuefm-8847,
    title={CoSENT: A more efficient sentence vector scheme than Sentence-BERT},
    author={Su Jianlin},
    year={2022},
    month={Jan},
    url={https://kexue.fm/archives/8847},
}