Jade-modernbert-ft

internally on leaderboard known as jade-ft-14-bert This is a sentence-transformers model finetuned from nomic-ai/modernbert-embed-base on the jade_embeddings_train_25.04.04 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: nomic-ai/modernbert-embed-base
• Maximum Sequence Length: 8192 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • jade_embeddings_train_25.04.04
• Language: en
• License: apache-2.0

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})
  (2): Normalize()
)

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("lwoollett/jade-ft-14-bert-static")
# Run inference
sentences = [
    'Which subclasses are associated with the JadeXMLCharacterData class?',
    '## JadeXMLCharacterData Class\n\nThe **JadeXMLCharacterData** class is the abstract superclass of character-based nodes in an XML document tree; that is, the text, **CDATA**, and comment nodes.\n\nFor details about the property defined in the **JadeXMLCharacterData** class, see "[JadeXMLCharacterData Property](jadexmlcharacterdata_property.htm)", in the following section.\n\n[JadeXMLNode](../jadexmlnode_class/jadexmlnode_class.htm)\n\n[JadeXMLCDATA](../jadexmlcdata_class/jadexmlcdata_class.htm), [JadeXMLComment](../jadexmlcomment_class/jadexmlcomment_class.htm), [JadeXMLText](../jadexmltext_class/jadexmltext_class.htm)',
    "### Minimizing the Working Set\n\nIn loops where there are multiple filters, apply the cheapest filters first and then the filters that reduce the working set the most. For example, consider the following code fragment, which finds sales of appliances in a specified city.\n\n```\nwhile iter.next(tran) do\r\n    if  tran.type = Type_Sale\r\n    and tran.myBranch.myLocation.city = targetCity\r\n    and tran.myProduct.isAppliance then\r\n        <do something with tran>\r\n    endif;\r\nendwhile;\n```\nIn this example, **tran.type** should be checked first, because it is the cheapest. The **tran** object must be fetched to evaluate all of the other conditions, so we may as well check the **type** attribute first. If we did the **isAppliance** check first, we would have to fetch all of the product objects for the transactions that were not sales. Regardless of how many transactions are sales and how many products are appliances, it will save time to check **tran.type** first.\n\nNow, assume that:\n\n- 80 percent of transactions are sales\n\n- 15 percent, on average, are likely to be in the target city\n\n- 90 percent of the products are appliances\n\nIt pays to check the city first, even though it means fetching the branch and location objects for the non‑appliance products. There are very few non‑appliance products, so the number of extra fetches is small. By contrast, checking for non‑appliance products for all other cities would result in a large number of extra fetches.\n\nIt doesn't matter if the filters are conditions of an [if](../../devref/ch1languageref/if_instruction.htm#if) instruction, multiple [if](../../devref/ch1languageref/if_instruction.htm#if) instructions, or multiple conditions in the [where](../../devref/ch1languageref/where_clause_optimization.htm#whereoptimization) clause of a [while](../../devref/ch1languageref/while_instruction.htm#while) statement; the end result is the same.\n\nThis code fragment example is simple and concise, to convey the concept. In the real world, each successive filter may be in another method, another class, or even another schema. It may take a bit of investigation to find all of the filters involved in a single loop.",
]
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

jade_embeddings_train_25.04.04

• Dataset: jade_embeddings_train_25.04.04
• Size: 10,217 training samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 8 tokens mean: 17.17 tokens max: 30 tokens	min: 27 tokens mean: 363.15 tokens max: 6303 tokens

• Samples:

  anchor	positive
  What is the format for defining a Byte constant in JADE?	##### Constant Definition Tips When defining a constant value, the value of a constant can be a simple literal value or an expression constructed using literals and other constants. For details about literal types, see "Literals", in Chapter - 1 of the Developer's Reference. You can define the value for a constant whose primitive type is not a specific literal format by using a typecast of a String literal or in the case of a Byte, a small Integer literal, as shown in the examples in the following table.
  How does the replaceFrom__ method handle case sensitivity?	#### replaceFrom__ ``` replaceFrom__(target: String;
  replacement: String;
  startIndex: Integer;
  bIgnoreCase: Boolean): String; ``` The replaceFrom__ method of the String primitive type replaces only the first occurrence of the substring specified in the target parameter with the substring specified in the replacement parameter, starting from the specified startIndex parameter. Case‑sensitivity is ignored if you set the value of the bIgnoreCase parameter to true. Set this parameter to false if you want the substring replacement to be case‑sensitive. This method raises exception 1413 (Index used in string operation is out of bounds) if the value specified in the startIndex parameter is less than 1 or it is greater than the length of the original string. In addition, it returns the original receiver String if the value specified in the target parameter has a length of zero (**...
  What does the global constant Ex_Continue do?	## Exceptions Category The global constants for exceptions are listed in the following table.

• Loss: CachedMultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim",
      "mini_batch_size": 32
  }
  

Evaluation Dataset

jade_embeddings_train_25.04.04

• Dataset: jade_embeddings_train_25.04.04
• Size: 1,136 evaluation samples
• Columns: anchor and positive
• Approximate statistics based on the first 1000 samples:

  	anchor	positive
  type	string	string
  details	min: 8 tokens mean: 17.07 tokens max: 41 tokens	min: 25 tokens mean: 365.93 tokens max: 3397 tokens

• Samples:

  anchor	positive
  What is the keyword list constant value for JADE_SYSTEMVARS?	### changeKeywords ``` changeKeywords(action: Integer;
  keywordList: Integer;
  keywords: String); ``` The changeKeywords method of the JadeTextEdit class modifies one or more of the current keyword lists. The keyword lists are used by the current language lexical analyzer to classify the tokens found in the text. For the Jade language, this includes keywords, class names, constant names, and so on. The value of the action parameter can be one of the JadeTextEdit class constants listed in the following table.	Class Constant
  What should you click to abandon the deletion of a report in JADE?	#### Delete Report Command Use the Delete Report command from the File menu to delete a report. To delete a report 1. Select the Delete Report command from the File menu. The Delete Report dialog, shown in the following image, is then displayed. 2. Select the report that you want to delete from the Report list box or enter the name in the Report name text box. 3. Filter the list of report names in the Reports list box in one or both of the following ways. - To display only those reports that contain that text in their report description, enter text in the Text contains text box. For example, only those reports that mention Pay in their description are displayed if you enter Pay, providing a refined selection list. - To display only those reports modified during a specified period, select a last modified period from the Last modified list box. For example, only those reports that were modified in...
  What types of objects can be set for the userGroupObject in JadeMultiWorkerTcpTransport?	#### userGroupObject Type: - Object The userGroupObject property of the JadeMultiWorkerTcpTransport class contains a reference to an object that you can associate with the transport group between event callbacks. You must set the value of this property to a shared transient or a persistent object, as it must be visible to other workers. The default value is null. To prevent an object leak, it is your responsibility to delete this object, if required, in your implementation of the closedEvent method in the receiver class.

• Loss: CachedMultipleNegativesRankingLoss with these parameters:

  {
      "scale": 20.0,
      "similarity_fct": "cos_sim",
      "mini_batch_size": 32
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 18
• per_device_eval_batch_size: 18
• num_train_epochs: 4
• warmup_ratio: 0.1
• bf16: True
• batch_sampler: no_duplicates

All Hyperparameters

Click to expand

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 18
• per_device_eval_batch_size: 18
• 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: 5e-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: 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}
• tp_size: 0
• 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: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	Validation Loss
0.1761	100	0.0851	0.0243
0.3521	200	0.0262	0.0211
0.5282	300	0.0275	0.0217
0.7042	400	0.0216	0.0256
0.8803	500	0.0283	0.0241
1.0563	600	0.0226	0.0195
1.2324	700	0.0113	0.0170
1.4085	800	0.0114	0.0204
1.5845	900	0.0165	0.0182
1.7606	1000	0.0129	0.0219
1.9366	1100	0.0126	0.0181
2.1127	1200	0.0069	0.0207
2.2887	1300	0.0045	0.0212
2.4648	1400	0.0046	0.0187
2.6408	1500	0.0056	0.0206
2.8169	1600	0.0084	0.0196
2.9930	1700	0.005	0.0214
3.1690	1800	0.0056	0.0202
3.3451	1900	0.0088	0.0190
3.5211	2000	0.0026	0.0202
3.6972	2100	0.0064	0.0205
3.8732	2200	0.006	0.0202

Framework Versions

• Python: 3.11.11
• Sentence Transformers: 4.0.2
• Transformers: 4.51.0
• PyTorch: 2.8.0.dev20250319+cu128
• Accelerate: 1.6.0
• Datasets: 3.5.0
• 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",
}

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}
}