tomaarsen/static-bert-uncased-gooaq

Static Embeddings with BERT uncased tokenizer finetuned on GooAQ pairs

This is a sentence-transformers model trained on the gooaq 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.

This model was trained using the train_script.py code.

Model Details

Model Description

• Model Type: Sentence Transformer
• Maximum Sequence Length: inf tokens
• Output Dimensionality: 1024 tokens
• Similarity Function: Cosine Similarity
• Training Dataset:
  • gooaq
• 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): StaticEmbedding(
    (embedding): EmbeddingBag(30522, 1024, mode='mean')
  )
)

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/static-bert-uncased-gooaq")
# Run inference
sentences = [
    "how to reverse a video on tiktok that's not yours?",
    '[\'Tap "Effects" at the bottom of your screen — it\\\'s an icon that looks like a clock. Open the Effects menu. ... \', \'At the end of the new list that appears, tap "Time." Select "Time" at the end. ... \', \'Select "Reverse" — you\\\'ll then see a preview of your new, reversed video appear on the screen.\']',
    'Relative age is the age of a rock layer (or the fossils it contains) compared to other layers. It can be determined by looking at the position of rock layers. Absolute age is the numeric age of a layer of rocks or fossils. Absolute age can be determined by using radiometric dating.',
]
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

• Dataset: gooaq-1024-dev
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6309
cosine_accuracy@3	0.8409
cosine_accuracy@5	0.8986
cosine_accuracy@10	0.9444
cosine_precision@1	0.6309
cosine_precision@3	0.2803
cosine_precision@5	0.1797
cosine_precision@10	0.0944
cosine_recall@1	0.6309
cosine_recall@3	0.8409
cosine_recall@5	0.8986
cosine_recall@10	0.9444
cosine_ndcg@10	0.7933
cosine_mrr@10	0.744
cosine_map@100	0.7466

Information Retrieval

• Dataset: gooaq-512-dev
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6271
cosine_accuracy@3	0.8366
cosine_accuracy@5	0.8946
cosine_accuracy@10	0.9431
cosine_precision@1	0.6271
cosine_precision@3	0.2789
cosine_precision@5	0.1789
cosine_precision@10	0.0943
cosine_recall@1	0.6271
cosine_recall@3	0.8366
cosine_recall@5	0.8946
cosine_recall@10	0.9431
cosine_ndcg@10	0.7905
cosine_mrr@10	0.7408
cosine_map@100	0.7434

Information Retrieval

• Dataset: gooaq-256-dev
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6192
cosine_accuracy@3	0.8235
cosine_accuracy@5	0.8866
cosine_accuracy@10	0.9364
cosine_precision@1	0.6192
cosine_precision@3	0.2745
cosine_precision@5	0.1773
cosine_precision@10	0.0936
cosine_recall@1	0.6192
cosine_recall@3	0.8235
cosine_recall@5	0.8866
cosine_recall@10	0.9364
cosine_ndcg@10	0.7821
cosine_mrr@10	0.7321
cosine_map@100	0.7349

Information Retrieval

• Dataset: gooaq-128-dev
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.5942
cosine_accuracy@3	0.804
cosine_accuracy@5	0.8721
cosine_accuracy@10	0.9249
cosine_precision@1	0.5942
cosine_precision@3	0.268
cosine_precision@5	0.1744
cosine_precision@10	0.0925
cosine_recall@1	0.5942
cosine_recall@3	0.804
cosine_recall@5	0.8721
cosine_recall@10	0.9249
cosine_ndcg@10	0.7628
cosine_mrr@10	0.7103
cosine_map@100	0.7134

Information Retrieval

• Dataset: gooaq-64-dev
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.556
cosine_accuracy@3	0.7553
cosine_accuracy@5	0.8267
cosine_accuracy@10	0.8945
cosine_precision@1	0.556
cosine_precision@3	0.2518
cosine_precision@5	0.1653
cosine_precision@10	0.0895
cosine_recall@1	0.556
cosine_recall@3	0.7553
cosine_recall@5	0.8267
cosine_recall@10	0.8945
cosine_ndcg@10	0.7246
cosine_mrr@10	0.6702
cosine_map@100	0.6743

Information Retrieval

• Dataset: gooaq-32-dev
• Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.4628
cosine_accuracy@3	0.6619
cosine_accuracy@5	0.7415
cosine_accuracy@10	0.8241
cosine_precision@1	0.4628
cosine_precision@3	0.2206
cosine_precision@5	0.1483
cosine_precision@10	0.0824
cosine_recall@1	0.4628
cosine_recall@3	0.6619
cosine_recall@5	0.7415
cosine_recall@10	0.8241
cosine_ndcg@10	0.6387
cosine_mrr@10	0.5798
cosine_map@100	0.5857

Training Details

Training Dataset

gooaq

• Dataset: gooaq at b089f72
• Size: 3,012,496 training samples
• Columns: question and answer
• Approximate statistics based on the first 1000 samples:

  	question	answer
  type	string	string
  details	min: 18 characters mean: 43.23 characters max: 96 characters	min: 55 characters mean: 253.36 characters max: 371 characters

• Samples:

  question	answer
  what is the difference between broilers and layers?	An egg laying poultry is called egger or layer whereas broilers are reared for obtaining meat. So a layer should be able to produce more number of large sized eggs, without growing too much. On the other hand, a broiler should yield more meat and hence should be able to grow well.
  what is the difference between chronological order and spatial order?	As a writer, you should always remember that unlike chronological order and the other organizational methods for data, spatial order does not take into account the time. Spatial order is primarily focused on the location. All it does is take into account the location of objects and not the time.
  is kamagra same as viagra?	Kamagra is thought to contain the same active ingredient as Viagra, sildenafil citrate. In theory, it should work in much the same way as Viagra, taking about 45 minutes to take effect, and lasting for around 4-6 hours. However, this will vary from person to person.

• Loss: MatryoshkaLoss with these parameters:

  {
      "loss": "MultipleNegativesRankingLoss",
      "matryoshka_dims": [
          1024,
          512,
          256,
          128,
          64,
          32
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Evaluation Dataset

gooaq

• Dataset: gooaq at b089f72
• Size: 3,012,496 evaluation samples
• Columns: question and answer
• Approximate statistics based on the first 1000 samples:

  	question	answer
  type	string	string
  details	min: 18 characters mean: 43.17 characters max: 98 characters	min: 51 characters mean: 254.12 characters max: 360 characters

• Samples:

  question	answer
  how do i program my directv remote with my tv?	['Press MENU on your remote.', 'Select Settings & Help > Settings > Remote Control > Program Remote.', 'Choose the device (TV, audio, DVD) you wish to program. ... ', 'Follow the on-screen prompts to complete programming.']
  are rodrigues fruit bats nocturnal?	Before its numbers were threatened by habitat destruction, storms, and hunting, some of those groups could number 500 or more members. Sunrise, sunset. Rodrigues fruit bats are most active at dawn, at dusk, and at night.
  why does your heart rate increase during exercise bbc bitesize?	During exercise there is an increase in physical activity and muscle cells respire more than they do when the body is at rest. The heart rate increases during exercise. The rate and depth of breathing increases - this makes sure that more oxygen is absorbed into the blood, and more carbon dioxide is removed from it.

• Loss: MatryoshkaLoss with these parameters:

  {
      "loss": "MultipleNegativesRankingLoss",
      "matryoshka_dims": [
          1024,
          512,
          256,
          128,
          64,
          32
      ],
      "matryoshka_weights": [
          1,
          1,
          1,
          1,
          1,
          1
      ],
      "n_dims_per_step": -1
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 2048
• per_device_eval_batch_size: 2048
• learning_rate: 0.2
• num_train_epochs: 1
• 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: 2048
• per_device_eval_batch_size: 2048
• 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: 0.2
• 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: 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}
• 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: False
• hub_always_push: False
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• 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
• dispatch_batches: None
• split_batches: 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
• eval_use_gather_object: False
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	Validation Loss	gooaq-1024-dev_cosine_map@100	gooaq-512-dev_cosine_map@100	gooaq-256-dev_cosine_map@100	gooaq-128-dev_cosine_map@100	gooaq-64-dev_cosine_map@100	gooaq-32-dev_cosine_map@100
0	0	-	-	0.2095	0.2010	0.1735	0.1381	0.0750	0.0331
0.0007	1	34.953	-	-	-	-	-	-	-
0.0682	100	16.2504	-	-	-	-	-	-	-
0.1363	200	5.9502	-	-	-	-	-	-	-
0.1704	250	-	1.6781	0.6791	0.6729	0.6619	0.6409	0.5904	0.4934
0.2045	300	4.8411	-	-	-	-	-	-	-
0.2727	400	4.336	-	-	-	-	-	-	-
0.3408	500	4.0484	1.3935	0.7104	0.7055	0.6968	0.6756	0.6322	0.5358
0.4090	600	3.8378	-	-	-	-	-	-	-
0.4772	700	3.6765	-	-	-	-	-	-	-
0.5112	750	-	1.2549	0.7246	0.7216	0.7133	0.6943	0.6482	0.5582
0.5453	800	3.5439	-	-	-	-	-	-	-
0.6135	900	3.4284	-	-	-	-	-	-	-
0.6817	1000	3.3576	1.1656	0.7359	0.7338	0.7252	0.7040	0.6604	0.5715
0.7498	1100	3.2456	-	-	-	-	-	-	-
0.8180	1200	3.2014	-	-	-	-	-	-	-
0.8521	1250	-	1.1133	0.7438	0.7398	0.7310	0.7099	0.6704	0.5796
0.8862	1300	3.1536	-	-	-	-	-	-	-
0.9543	1400	3.0696	-	-	-	-	-	-	-
1.0	1467	-	-	0.7466	0.7434	0.7349	0.7134	0.6743	0.5857

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.017 kWh
• Carbon Emitted: 0.006 kg of CO2
• Hours Used: 0.109 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: 3.2.0.dev0
• Transformers: 4.43.4
• PyTorch: 2.5.0.dev20240807+cu121
• Accelerate: 0.31.0
• Datasets: 2.20.0
• Tokenizers: 0.19.1

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",
}

MatryoshkaLoss

@misc{kusupati2024matryoshka,
    title={Matryoshka Representation Learning},
    author={Aditya Kusupati and Gantavya Bhatt and Aniket Rege and Matthew Wallingford and Aditya Sinha and Vivek Ramanujan and William Howard-Snyder and Kaifeng Chen and Sham Kakade and Prateek Jain and Ali Farhadi},
    year={2024},
    eprint={2205.13147},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

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}
}