benjamintli/modernbert-codesearchnet

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-codesearchnet")
# Run inference
queries = [
    "Split the data object along a given expression, in units.\n\n        Parameters\n        ----------\n        expression : int or str\n            The expression to split along. If given as an integer, the axis at that index\n            is used.\n        positions : number-type or 1D array-type\n            The position(s) to split at, in units.\n        units : str (optional)\n            The units of the given positions. Default is same, which assumes\n            input units are identical to first variable units.\n        parent : WrightTools.Collection (optional)\n            The parent collection in which to place the \u0027split\u0027 collection.\n            Default is a new Collection.\n        verbose : bool (optional)\n            Toggle talkback. Default is True.\n\n        Returns\n        -------\n        WrightTools.collection.Collection\n            A Collection of data objects.\n            The order of the objects is such that the axis points retain their original order.\n\n        See Also\n        --------\n        chop\n            Divide the dataset into its lower-dimensionality components.\n        collapse\n            Collapse the dataset along one axis.",
]
documents = [
    'def split(\n        self, expression, positions, *, units=None, parent=None, verbose=True\n    ) -> wt_collection.Collection:\n        """\n        Split the data object along a given expression, in units.\n\n        Parameters\n        ----------\n        expression : int or str\n            The expression to split along. If given as an integer, the axis at that index\n            is used.\n        positions : number-type or 1D array-type\n            The position(s) to split at, in units.\n        units : str (optional)\n            The units of the given positions. Default is same, which assumes\n            input units are identical to first variable units.\n        parent : WrightTools.Collection (optional)\n            The parent collection in which to place the \'split\' collection.\n            Default is a new Collection.\n        verbose : bool (optional)\n            Toggle talkback. Default is True.\n\n        Returns\n        -------\n        WrightTools.collection.Collection\n            A Collection of data objects.\n            The order of the objects is such that the axis points retain their original order.\n\n        See Also\n        --------\n        chop\n            Divide the dataset into its lower-dimensionality components.\n        collapse\n            Collapse the dataset along one axis.\n        """\n        # axis ------------------------------------------------------------------------------------\n        old_expr = self.axis_expressions\n        old_units = self.units\n        out = wt_collection.Collection(name="split", parent=parent)\n        if isinstance(expression, int):\n            if units is None:\n                units = self._axes[expression].units\n            expression = self._axes[expression].expression\n        elif isinstance(expression, str):\n            pass\n        else:\n            raise TypeError("expression: expected {int, str}, got %s" % type(expression))\n\n        self.transform(expression)\n        if units:\n            self.convert(units)\n\n        try:\n            positions = [-np.inf] + sorted(list(positions)) + [np.inf]\n        except TypeError:\n            positions = [-np.inf, positions, np.inf]\n\n        values = self._axes[0].full\n        masks = [(values >= lo) & (values < hi) for lo, hi in wt_kit.pairwise(positions)]\n        omasks = []\n        cuts = []\n        for mask in masks:\n            try:\n                omasks.append(wt_kit.mask_reduce(mask))\n                cuts.append([i == 1 for i in omasks[-1].shape])\n                # Ensure at least one axis is kept\n                if np.all(cuts[-1]):\n                    cuts[-1][0] = False\n            except ValueError:\n                omasks.append(None)\n                cuts.append(None)\n        for i in range(len(positions) - 1):\n            out.create_data("split%03i" % i)\n\n        for var in self.variables:\n            for i, (imask, omask, cut) in enumerate(zip(masks, omasks, cuts)):\n                if omask is None:\n                    # Zero length split\n                    continue\n                omask = wt_kit.enforce_mask_shape(omask, var.shape)\n                omask.shape = tuple([s for s, c in zip(omask.shape, cut) if not c])\n                out_arr = np.full(omask.shape, np.nan)\n                imask = wt_kit.enforce_mask_shape(imask, var.shape)\n                out_arr[omask] = var[:][imask]\n                out[i].create_variable(values=out_arr, **var.attrs)\n\n        for ch in self.channels:\n            for i, (imask, omask, cut) in enumerate(zip(masks, omasks, cuts)):\n                if omask is None:\n                    # Zero length split\n                    continue\n                omask = wt_kit.enforce_mask_shape(omask, ch.shape)\n                omask.shape = tuple([s for s, c in zip(omask.shape, cut) if not c])\n                out_arr = np.full(omask.shape, np.nan)\n                imask = wt_kit.enforce_mask_shape(imask, ch.shape)\n                out_arr[omask] = ch[:][imask]\n                out[i].create_channel(values=out_arr, **ch.attrs)\n\n        if verbose:\n            for d in out.values():\n                try:\n                    d.transform(expression)\n                except IndexError:\n                    continue\n\n            print("split data into {0} pieces along <{1}>:".format(len(positions) - 1, expression))\n            for i, (lo, hi) in enumerate(wt_kit.pairwise(positions)):\n                new_data = out[i]\n                if new_data.shape == ():\n                    print("  {0} : None".format(i))\n                else:\n                    new_axis = new_data.axes[0]\n                    print(\n                        "  {0} : {1:0.2f} to {2:0.2f} {3} {4}".format(\n                            i, lo, hi, new_axis.units, new_axis.shape\n                        )\n                    )\n\n        for d in out.values():\n            try:\n                d.transform(*old_expr)\n                keep = []\n                keep_units = []\n                for ax in d.axes:\n                    if ax.size > 1:\n                        keep.append(ax.expression)\n                        keep_units.append(ax.units)\n                    else:\n                        d.create_constant(ax.expression, verbose=False)\n                d.transform(*keep)\n                for ax, u in zip(d.axes, keep_units):\n                    ax.convert(u)\n            except IndexError:\n                continue\n            tempax = Axis(d, expression)\n            if all(\n                np.all(\n                    np.sum(~np.isnan(tempax.masked), axis=tuple(set(range(tempax.ndim)) - {j}))\n                    <= 1\n                )\n                for j in range(tempax.ndim)\n            ):\n                d.create_constant(expression, verbose=False)\n        self.transform(*old_expr)\n        for ax, u in zip(self.axes, old_units):\n            ax.convert(u)\n\n        return out',
    'def add_item(self, title, key, synonyms=None, description=None, img_url=None):\n        """Adds item to a list or carousel card.\n\n        A list must contain at least 2 items, each requiring a title and object key.\n\n        Arguments:\n            title {str} -- Name of the item object\n            key {str} -- Key refering to the item.\n                        This string will be used to send a query to your app if selected\n\n        Keyword Arguments:\n            synonyms {list} -- Words and phrases the user may send to select the item\n                              (default: {None})\n            description {str} -- A description of the item (default: {None})\n            img_url {str} -- URL of the image to represent the item (default: {None})\n        """\n        item = build_item(title, key, synonyms, description, img_url)\n        self._items.append(item)\n        return self',
    'def compare(a, b):\n    """Compares two timestamps.\n\n    ``a`` and ``b`` must be the same type, in addition to normal\n    representations of timestamps that order naturally, they can be rfc3339\n    formatted strings.\n\n    Args:\n      a (string|object): a timestamp\n      b (string|object): another timestamp\n\n    Returns:\n      int: -1 if a < b, 0 if a == b or 1 if a > b\n\n    Raises:\n      ValueError: if a or b are not the same type\n      ValueError: if a or b strings but not in valid rfc3339 format\n\n    """\n    a_is_text = isinstance(a, basestring)\n    b_is_text = isinstance(b, basestring)\n    if type(a) != type(b) and not (a_is_text and b_is_text):\n        _logger.error(u\'Cannot compare %s to %s, types differ %s!=%s\',\n                      a, b, type(a), type(b))\n        raise ValueError(u\'cannot compare inputs of differing types\')\n\n    if a_is_text:\n        a = from_rfc3339(a, with_nanos=True)\n        b = from_rfc3339(b, with_nanos=True)\n\n    if a < b:\n        return -1\n    elif a > b:\n        return 1\n    else:\n        return 0',
]
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([[0.9188, 0.1817, 0.1583]])

Evaluation

Metrics

Information Retrieval

• Dataset: eval
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.9481
cosine_accuracy@3	0.9703
cosine_accuracy@5	0.9752
cosine_accuracy@10	0.9807
cosine_precision@1	0.9481
cosine_precision@3	0.3234
cosine_precision@5	0.195
cosine_precision@10	0.0981
cosine_recall@1	0.9481
cosine_recall@3	0.9703
cosine_recall@5	0.9752
cosine_recall@10	0.9807
cosine_ndcg@10	0.9652
cosine_mrr@10	0.9602
cosine_map@100	0.9606

Training Details

Training Dataset

Unnamed Dataset

• Size: 369,762 training samples
• Columns: query and positive
• Approximate statistics based on the first 1000 samples:

  	query	positive
  type	string	string
  details	min: 3 tokens mean: 71.9 tokens max: 512 tokens	min: 37 tokens mean: 236.1 tokens max: 512 tokens

• Samples:

  query	positive
  Returns group object for datacenter root group. >>> clc.v2.Datacenter().RootGroup() >>> print _ WA1 Hardware	def RootGroup(self): """Returns group object for datacenter root group. >>> clc.v2.Datacenter().RootGroup() >>> print _ WA1 Hardware """ return(clc.v2.Group(id=self.root_group_id,alias=self.alias,session=self.session))
  Calculate the euclidean distance of all array positions in "matchArr". :param matchArr: a dictionary of numpy.arrays containing at least two entries that are treated as cartesian coordinates. :param tKey: #TODO: docstring :param mKey: #TODO: docstring :returns: #TODO: docstring {'eucDist': numpy.array([eucDistance, eucDistance, ...]), 'posPairs': numpy.array([[pos1, pos2], [pos1, pos2], ...]) }	def calcDistMatchArr(matchArr, tKey, mKey): """Calculate the euclidean distance of all array positions in "matchArr". :param matchArr: a dictionary of numpy.arrays containing at least two entries that are treated as cartesian coordinates. :param tKey: #TODO: docstring :param mKey: #TODO: docstring :returns: #TODO: docstring {'eucDist': numpy.array([eucDistance, eucDistance, ...]), 'posPairs': numpy.array([[pos1, pos2], [pos1, pos2], ...]) } """ #Calculate all sorted list of all eucledian feature distances matchArrSize = listvalues(matchArr)[0].size distInfo = {'posPairs': list(), 'eucDist': list()} _matrix = numpy.swapaxes(numpy.array([matchArr[tKey], matchArr[mKey]]), 0, 1) for pos1 in range(matchArrSize-1): for pos2 in range(pos1+1, matchArrSize): distInfo['posPairs'].append((pos1, pos2)) distInfo['posPairs'] = numpy.array(distInfo['posPairs']) distInfo['eucD...
  Format this verifier Returns: string: A formatted string	def format(self, indent_level, indent_size=4): """Format this verifier Returns: string: A formatted string """ name = self.format_name('Literal', indent_size) if self.long_desc is not None: name += '\n' name += self.wrap_lines('value: %s\n' % str(self._literal), 1, indent_size) return self.wrap_lines(name, indent_level, indent_size)

• 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: 19,462 evaluation samples
• Columns: query and positive
• Approximate statistics based on the first 1000 samples:

  	query	positive
  type	string	string
  details	min: 3 tokens mean: 71.05 tokens max: 512 tokens	min: 40 tokens mean: 236.22 tokens max: 512 tokens

• Samples:

  query	positive
  Create a new ParticipantInstance :param unicode attributes: An optional string metadata field you can use to store any data you wish. :param unicode twilio_address: The address of the Twilio phone number that the participant is in contact with. :param datetime date_created: The date that this resource was created. :param datetime date_updated: The date that this resource was last updated. :param unicode identity: A unique string identifier for the session participant as Chat User. :param unicode user_address: The address of the participant's device. :returns: Newly created ParticipantInstance :rtype: twilio.rest.messaging.v1.session.participant.ParticipantInstance	def create(self, attributes=values.unset, twilio_address=values.unset, date_created=values.unset, date_updated=values.unset, identity=values.unset, user_address=values.unset): """ Create a new ParticipantInstance :param unicode attributes: An optional string metadata field you can use to store any data you wish. :param unicode twilio_address: The address of the Twilio phone number that the participant is in contact with. :param datetime date_created: The date that this resource was created. :param datetime date_updated: The date that this resource was last updated. :param unicode identity: A unique string identifier for the session participant as Chat User. :param unicode user_address: The address of the participant's device. :returns: Newly created ParticipantInstance :rtype: twilio.rest.messaging.v1.session.participant.ParticipantInstance """ data = values.o...
  It returns absolute url defined by node related to this page	def get_absolute_url(self): """ It returns absolute url defined by node related to this page """ try: node = Node.objects.select_related().filter(page=self)[0] return node.get_absolute_url() except Exception, e: raise ValueError(u"Error in {0}.{1}: {2}".format(self.module, self.class.name, e)) return u""
  Return the current scaled font. :return: A new :class:ScaledFont object, wrapping an existing cairo object.	def get_scaled_font(self): """Return the current scaled font. :return: A new :class:ScaledFont object, wrapping an existing cairo object. """ return ScaledFont._from_pointer( cairo.cairo_get_scaled_font(self._pointer), incref=True)

• 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: 8192
• num_train_epochs: 1
• learning_rate: 2e-06
• warmup_steps: 0.1
• bf16: True
• eval_strategy: epoch
• per_device_eval_batch_size: 8192
• push_to_hub: True
• hub_model_id: modernbert-codesearchnet
• load_best_model_at_end: True
• dataloader_num_workers: 4
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• per_device_train_batch_size: 8192
• num_train_epochs: 1
• 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: 8192
• 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-codesearchnet
• 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
0.2174	10	0.9210	-	-
0.4348	20	0.6679	-	-
0.6522	30	0.5007	-	-
0.8696	40	0.4181	-	-
1.0	46	-	0.0328	0.9652

• 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}
}