benjamintli/modernbert-cosqa-hard-negatives

SentenceTransformer based on benjamintli/modernbert-cosqa

This is a sentence-transformers model finetuned from benjamintli/modernbert-cosqa on the cosqa-llm-filtered-hard-negatives 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: benjamintli/modernbert-cosqa
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • cosqa-llm-filtered-hard-negatives
• Language: en

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-hard-negatives")
# Run inference
sentences = [
    'python strftime miliseconds fixed width',
    'def fmt_duration(secs):\n    """Format a duration in seconds."""\n    return \' \'.join(fmt.human_duration(secs, 0, precision=2, short=True).strip().split())',
    'def seconds_to_hms(seconds):\n    """\n    Converts seconds float to \'hh:mm:ss.ssssss\' format.\n    """\n    hours = int(seconds / 3600.0)\n    minutes = int((seconds / 60.0) % 60.0)\n    secs = float(seconds % 60.0)\n    return "{0:02d}:{1:02d}:{2:02.6f}".format(hours, minutes, secs)',
]
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, 0.7076, 0.6960],
#         [0.7076, 1.0000, 0.7423],
#         [0.6960, 0.7423, 1.0000]])

Evaluation

Metrics

Information Retrieval

• Dataset: eval
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.5521
cosine_accuracy@3	0.8598
cosine_accuracy@5	0.9311
cosine_accuracy@10	0.9752
cosine_precision@1	0.5521
cosine_precision@3	0.2866
cosine_precision@5	0.1862
cosine_precision@10	0.0975
cosine_recall@1	0.5521
cosine_recall@3	0.8598
cosine_recall@5	0.9311
cosine_recall@10	0.9752
cosine_ndcg@10	0.7789
cosine_mrr@10	0.714
cosine_map@100	0.7154

Training Details

Training Dataset

cosqa-llm-filtered-hard-negatives

• Dataset: cosqa-llm-filtered-hard-negatives at 1585731
• Size: 19,963 training samples
• Columns: anchor, positive, and negative
• Approximate statistics based on the first 1000 samples:

  	anchor	positive	negative
  type	string	string	string
  details	min: 6 tokens mean: 9.57 tokens max: 22 tokens	min: 37 tokens mean: 88.3 tokens max: 512 tokens	min: 36 tokens mean: 94.31 tokens max: 512 tokens

• Samples:

  anchor	positive	negative
  python 2d array to dict	def to_dicts(recarray): """convert record array to a dictionaries""" for rec in recarray: yield dict(zip(recarray.dtype.names, rec.tolist()))	def multidict_to_dict(d): """ Turns a werkzeug.MultiDict or django.MultiValueDict into a dict with list values :param d: a MultiDict or MultiValueDict instance :return: a dict instance """ return dict((k, v[0] if len(v) == 1 else v) for k, v in iterlists(d))
  how to send dns request message in python	def _request_modify_dns_record(self, record): """Sends Modify_DNS_Record request""" return self._request_internal("Modify_DNS_Record", domain=self.domain, record=record)	def request(self, method, url, body=None, headers={}): """Send a complete request to the server.""" self._send_request(method, url, body, headers)
  how to cast string to uint8 in python	def b2u(string): """ bytes to unicode """ if (isinstance(string, bytes) or (PY2 and isinstance(string, str))): return string.decode('utf-8') return string	def to_bytes(s, encoding="utf-8"): """Convert a string to bytes.""" if isinstance(s, six.binary_type): return s if six.PY3: return bytes(s, encoding) return s.encode(encoding)

• 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

cosqa-llm-filtered-hard-negatives

• Dataset: cosqa-llm-filtered-hard-negatives at 1585731
• Size: 2,219 evaluation samples
• Columns: anchor, positive, and negative
• Approximate statistics based on the first 1000 samples:

  	anchor	positive	negative
  type	string	string	string
  details	min: 6 tokens mean: 9.65 tokens max: 21 tokens	min: 35 tokens mean: 88.87 tokens max: 512 tokens	min: 37 tokens mean: 94.56 tokens max: 512 tokens

• Samples:

  anchor	positive	negative
  way to change the string "python" to have all uppercase letters	def uppercase_chars(string: any) -> str: """Return all (and only) the uppercase chars in the given string.""" return ''.join([c if c.isupper() else '' for c in str(string)])	def to_capitalized_camel_case(snake_case_string): """ Convert a string from snake case to camel case with the first letter capitalized. For example, "some_var" would become "SomeVar". :param snake_case_string: Snake-cased string to convert to camel case. :returns: Camel-cased version of snake_case_string. """ parts = snake_case_string.split('_') return ''.join([i.title() for i in parts])
  how to make intercept zero in python	def prox_zero(X, step): """Proximal operator to project onto zero """ return np.zeros(X.shape, dtype=X.dtype)	def _adjust_offset(self, real_wave_mfcc, algo_parameters): """ OFFSET """ self.log(u"Called _adjust_offset") self._apply_offset(offset=algo_parameters[0])
  stop running function and passing to other variable python	def stop(self) -> None: """Stops the analysis as soon as possible.""" if self._stop and not self._posted_kork: self._stop() self._stop = None	def stop(self, dummy_signum=None, dummy_frame=None): """ Shutdown process (this method is also a signal handler) """ logging.info('Shutting down ...') self.socket.close() sys.exit(0)

• 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-hard-negatives
• 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-hard-negatives
• 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
0.5	10	1.5380	-	-
1.0	20	1.4167	0.9702	0.7440
1.5	30	1.4515	-	-
2.0	40	1.3789	0.9269	0.7499
2.5	50	1.3920	-	-
3.0	60	1.2849	0.8898	0.7581
3.5	70	1.3585	-	-
4.0	80	1.2197	0.8572	0.7653
4.5	90	1.2825	-	-
5.0	100	1.2078	0.8350	0.7686
5.5	110	1.2496	-	-
6.0	120	1.1569	0.8104	0.7720
6.5	130	1.2119	-	-
7.0	140	1.1278	0.7952	0.7754
7.5	150	1.1812	-	-
8.0	160	1.1018	0.7835	0.7770
8.5	170	1.1696	-	-
9.0	180	1.0972	0.7788	0.7786
9.5	190	1.1655	-	-
10.0	200	1.0796	0.7755	0.7789

• 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}
}