tomaarsen/mxbai-embed-large-v1-static-queries-nq

mxbai-embed-large-v1 with static query embeddings trained on Natural Questions pairs

This is a sentence-transformers model trained on the natural-questions dataset. It maps sentences & paragraphs to a 1024-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
• Maximum Sequence Length: inf tokens
• Output Dimensionality: 1024 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • natural-questions
• 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): Router(
    (query_0_StaticEmbedding): StaticEmbedding(
      (embedding): EmbeddingBag(30522, 1024, mode='mean')
    )
    (document_0_Transformer): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: BertModel 
    (document_1_Pooling): Pooling({'word_embedding_dimension': 1024, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  )
  (1): 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("tomaarsen/mxbai-embed-large-v1-static-queries-nq")
# Run inference
sentences = [
    'where does the last name francisco come from',
    'Francisco Francisco is the Spanish and Portuguese form of the masculine given name Franciscus (corresponding to English Francis).',
    'Book of Esther The Book of Esther, also known in Hebrew as "the Scroll" (Megillah), is a book in the third section (Ketuvim, "Writings") of the Jewish Tanakh (the Hebrew Bible) and in the Christian Old Testament. It is one of the five Scrolls (Megillot) in the Hebrew Bible. It relates the story of a Hebrew woman in Persia, born as Hadassah but known as Esther, who becomes queen of Persia and thwarts a genocide of her people. The story forms the core of the Jewish festival of Purim, during which it is read aloud twice: once in the evening and again the following morning. The books of Esther and Song of Songs are the only books in the Hebrew Bible that do not explicitly mention God.[2]',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 1024]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

Evaluation

Metrics

Information Retrieval

• Datasets: NanoMSMARCO, NanoNFCorpus and NanoNQ
• Evaluated with InformationRetrievalEvaluator

Metric	NanoMSMARCO	NanoNFCorpus	NanoNQ
cosine_accuracy@1	0.16	0.08	0.2
cosine_accuracy@3	0.24	0.2	0.42
cosine_accuracy@5	0.28	0.28	0.48
cosine_accuracy@10	0.34	0.36	0.56
cosine_precision@1	0.16	0.08	0.2
cosine_precision@3	0.08	0.0933	0.14
cosine_precision@5	0.056	0.092	0.1
cosine_precision@10	0.034	0.084	0.058
cosine_recall@1	0.16	0.0036	0.18
cosine_recall@3	0.24	0.0078	0.38
cosine_recall@5	0.28	0.0327	0.45
cosine_recall@10	0.34	0.0428	0.53
cosine_ndcg@10	0.2373	0.0935	0.3614
cosine_mrr@10	0.2058	0.1664	0.323
cosine_map@100	0.2267	0.0335	0.318

Nano BEIR

• Dataset: NanoBEIR_mean
• Evaluated with NanoBEIREvaluator with these parameters:

  {
      "dataset_names": [
          "msmarco",
          "nfcorpus",
          "nq"
      ]
  }
  

Metric	Value
cosine_accuracy@1	0.1467
cosine_accuracy@3	0.2867
cosine_accuracy@5	0.3467
cosine_accuracy@10	0.42
cosine_precision@1	0.1467
cosine_precision@3	0.1044
cosine_precision@5	0.0827
cosine_precision@10	0.0587
cosine_recall@1	0.1145
cosine_recall@3	0.2093
cosine_recall@5	0.2542
cosine_recall@10	0.3043
cosine_ndcg@10	0.2307
cosine_mrr@10	0.2317
cosine_map@100	0.1927

Training Details

Training Dataset

natural-questions

• Dataset: natural-questions at f9e894e
• Size: 99,231 training samples
• Columns: query and answer
• Approximate statistics based on the first 1000 samples:

  	query	answer
  type	string	string
  details	min: 10 tokens mean: 11.74 tokens max: 24 tokens	min: 15 tokens mean: 137.2 tokens max: 508 tokens

• Samples:

  query	answer
  who is required to report according to the hmda	Home Mortgage Disclosure Act US financial institutions must report HMDA data to their regulator if they meet certain criteria, such as having assets above a specific threshold. The criteria is different for depository and non-depository institutions and are available on the FFIEC website.[4] In 2012, there were 7,400 institutions that reported a total of 18.7 million HMDA records.[5]
  what is the definition of endoplasmic reticulum in biology	Endoplasmic reticulum The endoplasmic reticulum (ER) is a type of organelle in eukaryotic cells that forms an interconnected network of flattened, membrane-enclosed sacs or tube-like structures known as cisternae. The membranes of the ER are continuous with the outer nuclear membrane. The endoplasmic reticulum occurs in most types of eukaryotic cells, but is absent from red blood cells and spermatozoa. There are two types of endoplasmic reticulum: rough and smooth. The outer (cytosolic) face of the rough endoplasmic reticulum is studded with ribosomes that are the sites of protein synthesis. The rough endoplasmic reticulum is especially prominent in cells such as hepatocytes. The smooth endoplasmic reticulum lacks ribosomes and functions in lipid manufacture and metabolism, the production of steroid hormones, and detoxification.[1] The smooth ER is especially abundant in mammalian liver and gonad cells. The lacy membranes of the endoplasmic reticulum were first seen in 1945 using elect...
  what does the ski mean in polish names	Polish name Since the High Middle Ages, Polish-sounding surnames ending with the masculine -ski suffix, including -cki and -dzki, and the corresponding feminine suffix -ska/-cka/-dzka were associated with the nobility (Polish szlachta), which alone, in the early years, had such suffix distinctions.[1] They are widely popular today.

• Loss: MultipleNegativesRankingLoss with these parameters:

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

Evaluation Dataset

natural-questions

• Dataset: natural-questions at f9e894e
• Size: 1,000 evaluation samples
• Columns: query and answer
• Approximate statistics based on the first 1000 samples:

  	query	answer
  type	string	string
  details	min: 10 tokens mean: 11.78 tokens max: 22 tokens	min: 11 tokens mean: 135.64 tokens max: 512 tokens

• Samples:

  query	answer
  difference between russian blue and british blue cat	Russian Blue The coat is known as a "double coat", with the undercoat being soft, downy and equal in length to the guard hairs, which are an even blue with silver tips. However, the tail may have a few very dull, almost unnoticeable stripes. The coat is described as thick, plush and soft to the touch. The feeling is softer than the softest silk. The silver tips give the coat a shimmering appearance. Its eyes are almost always a dark and vivid green. Any white patches of fur or yellow eyes in adulthood are seen as flaws in show cats.[3] Russian Blues should not be confused with British Blues (which are not a distinct breed, but rather a British Shorthair with a blue coat as the British Shorthair breed itself comes in a wide variety of colors and patterns), nor the Chartreux or Korat which are two other naturally occurring breeds of blue cats, although they have similar traits.
  who played the little girl on mrs doubtfire	Mara Wilson Mara Elizabeth Wilson[2] (born July 24, 1987) is an American writer and former child actress. She is known for playing Natalie Hillard in Mrs. Doubtfire (1993), Susan Walker in Miracle on 34th Street (1994), Matilda Wormwood in Matilda (1996) and Lily Stone in Thomas and the Magic Railroad (2000). Since retiring from film acting, Wilson has focused on writing.
  what year did the movie the sound of music come out	The Sound of Music (film) The film was released on March 2, 1965 in the United States, initially as a limited roadshow theatrical release. Although critical response to the film was widely mixed, the film was a major commercial success, becoming the number one box office movie after four weeks, and the highest-grossing film of 1965. By November 1966, The Sound of Music had become the highest-grossing film of all-time—surpassing Gone with the Wind—and held that distinction for five years. The film was just as popular throughout the world, breaking previous box-office records in twenty-nine countries. Following an initial theatrical release that lasted four and a half years, and two successful re-releases, the film sold 283 million admissions worldwide and earned a total worldwide gross of $286,000,000.

• Loss: MultipleNegativesRankingLoss with these parameters:

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

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: 1
• warmup_ratio: 0.1
• seed: 12
• bf16: True
• batch_sampler: no_duplicates
• router_mapping: {'query': 'query', 'answer': 'document'}
• learning_rate_mapping: {'StaticEmbedding\.embedding': 0.2}

All Hyperparameters

Click to expand

• 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: 1
• 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: 12
• 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}
• 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
• router_mapping: {'query': 'query', 'answer': 'document'}
• learning_rate_mapping: {'StaticEmbedding\.embedding': 0.2}

Training Logs

Epoch	Step	Training Loss	Validation Loss	NanoMSMARCO_cosine_ndcg@10	NanoNFCorpus_cosine_ndcg@10	NanoNQ_cosine_ndcg@10	NanoBEIR_mean_cosine_ndcg@10
-1	-1	-	-	0.0	0.0195	0.0	0.0065
0.0002	1	2.9289	-	-	-	-	-
0.0322	200	2.7795	-	-	-	-	-
0.0645	400	2.137	-	-	-	-	-
0.0967	600	1.5211	1.1857	0.0440	0.0317	0.1087	0.0615
0.1290	800	1.0909	-	-	-	-	-
0.1612	1000	0.8669	-	-	-	-	-
0.1935	1200	0.7003	0.5961	0.0990	0.0503	0.2530	0.1341
0.2257	1400	0.5979	-	-	-	-	-
0.2580	1600	0.5242	-	-	-	-	-
0.2902	1800	0.4695	0.4039	0.1633	0.0596	0.2845	0.1691
0.3225	2000	0.4223	-	-	-	-	-
0.3547	2200	0.4145	-	-	-	-	-
0.3870	2400	0.3736	0.3128	0.1958	0.0717	0.2990	0.1888
0.4192	2600	0.3325	-	-	-	-	-
0.4515	2800	0.3172	-	-	-	-	-
0.4837	3000	0.2966	0.2590	0.1948	0.0744	0.3058	0.1917
0.5160	3200	0.2741	-	-	-	-	-
0.5482	3400	0.281	-	-	-	-	-
0.5805	3600	0.2533	0.2269	0.2113	0.0805	0.3407	0.2108
0.6127	3800	0.248	-	-	-	-	-
0.6450	4000	0.2402	-	-	-	-	-
0.6772	4200	0.2267	0.2044	0.2188	0.0810	0.3396	0.2131
0.7094	4400	0.2172	-	-	-	-	-
0.7417	4600	0.2277	-	-	-	-	-
0.7739	4800	0.2047	0.1905	0.2276	0.0893	0.3352	0.2173
0.8062	5000	0.2011	-	-	-	-	-
0.8384	5200	0.198	-	-	-	-	-
0.8707	5400	0.2025	0.1826	0.2439	0.0939	0.3443	0.2274
0.9029	5600	0.2018	-	-	-	-	-
0.9352	5800	0.1896	-	-	-	-	-
0.9674	6000	0.1973	0.1783	0.2373	0.0919	0.3614	0.2302
0.9997	6200	0.1924	-	-	-	-	-
-1	-1	-	-	0.2373	0.0935	0.3614	0.2307

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.137 kWh
• Carbon Emitted: 0.053 kg of CO2
• Hours Used: 0.344 hours

Training Hardware

• On Cloud: No
• GPU Model: 1 x NVIDIA GeForce RTX 3090
• CPU Model: 13th Gen Intel(R) Core(TM) i7-13700K
• RAM Size: 31.78 GB

Framework Versions

• Python: 3.11.6
• Sentence Transformers: 4.2.0.dev0
• Transformers: 4.52.3
• PyTorch: 2.6.0+cu124
• Accelerate: 1.5.1
• Datasets: 2.21.0
• 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",
}

MultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}