SentenceTransformer based on benjamintli/modernbert-cosqa

This is a sentence-transformers model finetuned from benjamintli/modernbert-cosqa. 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: benjamintli/modernbert-cosqa
• Maximum Sequence Length: 512 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': 512, 'do_lower_case': False, 'architecture': 'OptimizedModule'})
  (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("modernbert-cosqa")
# Run inference
queries = [
    "first duplicate element in list in python",
]
documents = [
    'def purge_duplicates(list_in):\n    """Remove duplicates from list while preserving order.\n\n    Parameters\n    ----------\n    list_in: Iterable\n\n    Returns\n    -------\n    list\n        List of first occurences in order\n    """\n    _list = []\n    for item in list_in:\n        if item not in _list:\n            _list.append(item)\n    return _list',
    'def getRect(self):\n\t\t"""\n\t\tReturns the window bounds as a tuple of (x,y,w,h)\n\t\t"""\n\t\treturn (self.x, self.y, self.w, self.h)',
    'def python_mime(fn):\n    """\n    Decorator, which adds correct MIME type for python source to the decorated\n    bottle API function.\n    """\n    @wraps(fn)\n    def python_mime_decorator(*args, **kwargs):\n        response.content_type = "text/x-python"\n\n        return fn(*args, **kwargs)\n\n    return python_mime_decorator',
]
query_embeddings = model.encode_query(queries)
document_embeddings = model.encode_document(documents)
print(query_embeddings.shape, document_embeddings.shape)
# [1, 768] [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(query_embeddings, document_embeddings)
print(similarities)
# tensor([[ 0.5986, -0.0006, -0.0122]])

Evaluation

Metrics

Information Retrieval

• Dataset: eval
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6197
cosine_accuracy@3	0.8847
cosine_accuracy@5	0.939
cosine_accuracy@10	0.9778
cosine_precision@1	0.6197
cosine_precision@3	0.2949
cosine_precision@5	0.1878
cosine_precision@10	0.0978
cosine_recall@1	0.6197
cosine_recall@3	0.8847
cosine_recall@5	0.939
cosine_recall@10	0.9778
cosine_ndcg@10	0.8125
cosine_mrr@10	0.7577
cosine_map@100	0.7588

Training Details

Training Dataset

Unnamed Dataset

• Size: 8,118 training samples
• Columns: query and positive
• Approximate statistics based on the first 1000 samples:

  	query	positive
  type	string	string
  details	min: 6 tokens mean: 9.3 tokens max: 23 tokens	min: 35 tokens mean: 85.05 tokens max: 512 tokens

• Samples:

  query	positive
  python code for opening geojson file	def _loadfilepath(self, filepath, **kwargs): """This loads a geojson file into a geojson python dictionary using the json module. Note: to load with a different text encoding use the encoding argument. """ with open(filepath, "r") as f: data = json.load(f, **kwargs) return data
  python 3 none compare with int	def is_natural(x): """A non-negative integer.""" try: is_integer = int(x) == x except (TypeError, ValueError): return False return is_integer and x >= 0
  design db memory cache python	def refresh(self, document): """ Load a new copy of a document from the database. does not replace the old one """ try: old_cache_size = self.cache_size self.cache_size = 0 obj = self.query(type(document)).filter_by(mongo_id=document.mongo_id).one() finally: self.cache_size = old_cache_size self.cache_write(obj) return obj

• Loss: CachedMultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "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

Unnamed Dataset

• Size: 902 evaluation samples
• Columns: query and positive
• Approximate statistics based on the first 902 samples:

  	query	positive
  type	string	string
  details	min: 6 tokens mean: 9.24 tokens max: 22 tokens	min: 38 tokens mean: 86.55 tokens max: 332 tokens

• Samples:

  query	positive
  how to remove masked items in python array	def ma(self): """Represent data as a masked array. The array is returned with column-first indexing, i.e. for a data file with columns X Y1 Y2 Y3 ... the array a will be a[0] = X, a[1] = Y1, ... . inf and nan are filtered via :func:numpy.isfinite. """ a = self.array return numpy.ma.MaskedArray(a, mask=numpy.logical_not(numpy.isfinite(a)))
  python deepcopy basic type	def deepcopy(self, memo): """Improve deepcopy speed.""" return type(self)(value=self._value, enum_ref=self.enum_ref)
  python number of non nan rows in a row	def count_rows_with_nans(X): """Count the number of rows in 2D arrays that contain any nan values.""" if X.ndim == 2: return np.where(np.isnan(X).sum(axis=1) != 0, 1, 0).sum()

• Loss: CachedMultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "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: 1024
• num_train_epochs: 10
• learning_rate: 2e-06
• warmup_steps: 0.1
• bf16: True
• eval_strategy: epoch
• per_device_eval_batch_size: 1024
• push_to_hub: True
• hub_model_id: modernbert-cosqa
• load_best_model_at_end: True
• dataloader_num_workers: 4
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• per_device_train_batch_size: 1024
• num_train_epochs: 10
• max_steps: -1
• learning_rate: 2e-06
• lr_scheduler_type: linear
• lr_scheduler_kwargs: None
• warmup_steps: 0.1
• 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: 1
• 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: epoch
• per_device_eval_batch_size: 1024
• 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: True
• hub_private_repo: None
• hub_model_id: modernbert-cosqa
• 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: 4
• 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: no_duplicates
• multi_dataset_batch_sampler: proportional
• router_mapping: {}
• learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	Validation Loss	eval_cosine_ndcg@10
1.0	8	-	0.3550	0.8071
1.25	10	1.0218	-	-
2.0	16	-	0.3508	0.8110
2.5	20	0.9890	-	-
3.0	24	-	0.3466	0.8131
3.75	30	0.9778	-	-
4.0	32	-	0.3439	0.8136
5.0	40	0.9507	0.3417	0.8148
6.0	48	-	0.3404	0.8120
6.25	50	0.9429	-	-
7.0	56	-	0.3387	0.8131
7.5	60	0.9267	-	-
8.0	64	-	0.3378	0.8127
8.75	70	0.9396	-	-
9.0	72	-	0.3370	0.8106
10.0	80	0.9099	0.3366	0.8125

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.12.12
• Sentence Transformers: 5.3.0
• Transformers: 5.3.0
• PyTorch: 2.10.0+cu128
• Accelerate: 1.13.0
• Datasets: 4.8.2
• Tokenizers: 0.22.2

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}
}