ModernBERT-small

SentenceTransformer (Legacy)

UPDATE: Consider using johnnyboycurtis/ModernBERT-small-v2 a much more performant model.

Warning

This model was an early exploration into creating a Wide model.

⚠️ Legacy Status: NOT RECOMMENDED.

This initial implementation suffered from suboptimal architectural scaling decisions made during the initialization phase, particularly concerning the feed-forward network capacity relative to the depth.

👉 Recommended Successor: For superior performance, speed, and architectural coherence, please use the improved version: johnnyboycurtis/ModernBERT-small-v2. The successor model addresses these limitations via a more sophisticated Guided Weight Initialization (GUIDE) technique and specialized Knowledge Distillation tuning.

Model Details

This is a shallow model with wide layers. NOT RECOMMENDED for production. This was my first attempt of training a ModernBERT model from scratch. The wide component is a mistake on my part due to lack of understanding of the gegelu design.

ModernBERT-small-1.5 will address the limitations of this design.

small_modernbert_config = ModernBertConfig(
    hidden_size=384,                 # A common dimension for small embedding models
    num_hidden_layers=12,               # Significantly fewer layers than the base's 22
    num_attention_heads=6,             # Must be a divisor of hidden_size
    intermediate_size=1536,            # 4 * hidden_size -- VERY WIDE!!
    max_position_embeddings=1024,       # Max sequence length for the model; originally 8192
)

model = ModernBertModel(modernbert_small_config)

Model Description

• Model Type: Sentence Transformer
• Base model: Custom-trained ModernBERT-Small (trained from scratch)
• Architecture: ModernBERT-Small
• Maximum Sequence Length: 1024 tokens
• Output Dimensionality: 384 dimensions
• Similarity Function: Cosine Similarity
• Language: en
• License: MIT

Model Sources

• Repository: ModernBERT Training Scripts
• Documentation: Sentence Transformers Documentation
• Hugging Face: Sentence Transformers on Hugging Face

Training Procedure

This model was developed using a sophisticated, multi-stage "curriculum learning" approach to build a deep semantic understanding. The training scripts are available in the linked repository.

Stage 1: Foundational Contrastive Training

The model was first trained on a large, diverse collection of over 1 million triplets from three different datasets. This stage taught the model a broad, foundational understanding of language, relevance, and logical relationships.

• Datasets:
  • sentence-transformers/all-nli
  • sentence-transformers/trivia-qa-triplet
  • sentence-transformers/msmarco-msmarco-distilbert-base-v3
• Loss Function: MultipleNegativesRankingLoss

Stage 2: Advanced Knowledge Distillation

The foundational model was then refined by having it mimic a state-of-the-art teacher model (BAAI/bge-base-en-v1.5). This stage transferred the nuanced knowledge of the expert teacher to our more efficient student model.

• Teacher Model: BAAI/bge-base-en-v1.5
• Loss Function: DistillKLDivLoss

Stage 3: Task-Specific Fine-Tuning

As a final "calibration" step, the best distilled model was fine-tuned directly on the Semantic Textual Similarity (STS) benchmark. This specializes the model for tasks requiring precise similarity scores.

• Dataset: sentence-transformers/stsb
• Loss Function: CosineSimilarityLoss

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 1024, 'do_lower_case': False, 'architecture': 'ModernBertModel'})
  (1): Pooling({'word_embedding_dimension': 384, '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("johnnyboycurtis/ModernBERT-small")
# Run inference
sentences = [
    'A girl is eating a cupcake.',
    'A woman is eating a cupcake.',
    'Zebras are socializing.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 384]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities)
# tensor([[1.0000, 0.8201, 0.1449],
#         [0.8201, 1.0000, 0.1839],
#         [0.1449, 0.1839, 1.0000]])

<!-- ### Direct Usage (Transformers) <details><summary>Click to see the direct usage in Transformers</summary> </details> --> <!-- ### Downstream Usage (Sentence Transformers) You can finetune this model on your own dataset. <details><summary>Click to expand</summary> </details> --> <!-- ### Out-of-Scope Use *List how the model may foreseeably be misused and address what users ought not to do with the model.* -->

Evaluation

Metrics

Semantic Similarity

• Datasets: sts-dev and sts-test
• Evaluated with <code>EmbeddingSimilarityEvaluator</code>

<!-- ## Bias, Risks and Limitations *What are the known or foreseeable issues stemming from this model? You could also flag here known failure cases or weaknesses of the model.* --> <!-- ### Recommendations *What are recommendations with respect to the foreseeable issues? For example, filtering explicit content.* -->

Training Details

Training Dataset

stsb

• Dataset: stsb at ab7a5ac
• Size: 5,749 training samples
• Columns: <code>sentence1</code>, <code>sentence2</code>, and <code>score</code>
• Approximate statistics based on the first 1000 samples:

  	sentence1	sentence2	score
  type	string	string	float
  details	<ul><li>min: 6 tokens</li><li>mean: 10.16 tokens</li><li>max: 28 tokens</li></ul>	<ul><li>min: 6 tokens</li><li>mean: 10.12 tokens</li><li>max: 25 tokens</li></ul>	<ul><li>min: 0.0</li><li>mean: 0.45</li><li>max: 1.0</li></ul>

• Samples:

  sentence1	sentence2	score
  <code>A plane is taking off.</code>	<code>An air plane is taking off.</code>	<code>1.0</code>
  <code>A man is playing a large flute.</code>	<code>A man is playing a flute.</code>	<code>0.76</code>
  <code>A man is spreading shreded cheese on a pizza.</code>	<code>A man is spreading shredded cheese on an uncooked pizza.</code>	<code>0.76</code>

• Loss: <code>CosineSimilarityLoss</code> with these parameters:

  {
      "loss_fct": "torch.nn.modules.loss.MSELoss"
  }
  

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: 4
• warmup_ratio: 0.1
• bf16: True
• load_best_model_at_end: True

All Hyperparameters

<details><summary>Click to expand</summary>

• 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: 4
• 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: 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: 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}
• 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
• 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: batch_sampler
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

</details>

Training Logs

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.13.4
• Sentence Transformers: 5.0.0
• Transformers: 4.52.4
• PyTorch: 2.7.1+cu128
• Accelerate: 1.7.0
• Datasets: 3.6.0
• Tokenizers: 0.21.1

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",
}

<!-- ## Glossary *Clearly define terms in order to be accessible across audiences.* --> <!-- ## Model Card Authors *Lists the people who create the model card, providing recognition and accountability for the detailed work that goes into its construction.* --> <!-- ## Model Card Contact *Provides a way for people who have updates to the Model Card, suggestions, or questions, to contact the Model Card authors.* -->