tomaarsen/mpnet-base-nq-cgist-triplet-mask-initial

MPNet base trained on Natural Questions pairs

This is a sentence-transformers model finetuned from microsoft/mpnet-base on the gooaq-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: microsoft/mpnet-base
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 dimensions
• Similarity Function: Cosine Similarity
• Training Dataset:
  • gooaq-hard-negatives
• 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': 512, 'do_lower_case': False}) with Transformer model: MPNetModel 
  (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("tomaarsen/mpnet-base-nq-cgist-triplet-mask")
# Run inference
sentences = [
    'what energy is released when coal is burned?',
    'When coal is burned, it reacts with the oxygen in the air. This chemical reaction converts the stored solar energy into thermal energy, which is released as heat. But it also produces carbon dioxide and methane.',
    'When coal is burned it releases a number of airborne toxins and pollutants. They include mercury, lead, sulfur dioxide, nitrogen oxides, particulates, and various other heavy metals.',
]
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]

Evaluation

Metrics

Information Retrieval

• Datasets: NanoClimateFEVER, NanoDBPedia, NanoFEVER, NanoFiQA2018, NanoHotpotQA, NanoMSMARCO, NanoNFCorpus, NanoNQ, NanoQuoraRetrieval, NanoSCIDOCS, NanoArguAna, NanoSciFact and NanoTouche2020
• Evaluated with InformationRetrievalEvaluator

Metric	NanoClimateFEVER	NanoDBPedia	NanoFEVER	NanoFiQA2018	NanoHotpotQA	NanoMSMARCO	NanoNFCorpus	NanoNQ	NanoQuoraRetrieval	NanoSCIDOCS	NanoArguAna	NanoSciFact	NanoTouche2020
cosine_accuracy@1	0.22	0.44	0.38	0.26	0.34	0.12	0.32	0.16	0.82	0.34	0.18	0.36	0.5306
cosine_accuracy@3	0.44	0.66	0.54	0.5	0.58	0.34	0.42	0.36	0.9	0.48	0.54	0.46	0.7347
cosine_accuracy@5	0.5	0.74	0.58	0.52	0.64	0.54	0.44	0.46	0.92	0.54	0.62	0.48	0.8367
cosine_accuracy@10	0.72	0.82	0.68	0.58	0.74	0.66	0.48	0.56	0.96	0.66	0.84	0.6	0.9592
cosine_precision@1	0.22	0.44	0.38	0.26	0.34	0.12	0.32	0.16	0.82	0.34	0.18	0.36	0.5306
cosine_precision@3	0.1667	0.3867	0.18	0.22	0.2133	0.1133	0.22	0.12	0.3733	0.2533	0.18	0.1667	0.4558
cosine_precision@5	0.116	0.372	0.12	0.16	0.144	0.108	0.188	0.092	0.244	0.216	0.124	0.104	0.4082
cosine_precision@10	0.092	0.346	0.07	0.098	0.094	0.066	0.138	0.058	0.134	0.148	0.084	0.066	0.349
cosine_recall@1	0.0933	0.0304	0.37	0.1343	0.17	0.12	0.0122	0.15	0.7207	0.0707	0.18	0.325	0.0388
cosine_recall@3	0.195	0.0762	0.52	0.3227	0.32	0.34	0.0207	0.34	0.862	0.1577	0.54	0.44	0.1011
cosine_recall@5	0.2267	0.12	0.57	0.3654	0.36	0.54	0.0267	0.43	0.8993	0.2227	0.62	0.46	0.1436
cosine_recall@10	0.3673	0.2194	0.66	0.4307	0.47	0.66	0.0374	0.53	0.9567	0.3057	0.84	0.585	0.2368
cosine_ndcg@10	0.2733	0.3883	0.5162	0.3408	0.3824	0.3693	0.1698	0.3369	0.8834	0.2912	0.495	0.4563	0.4021
cosine_mrr@10	0.3627	0.5594	0.4761	0.3757	0.4719	0.278	0.3737	0.2855	0.8639	0.4295	0.3864	0.4282	0.6537
cosine_map@100	0.2031	0.2782	0.477	0.283	0.3071	0.2903	0.0485	0.289	0.8576	0.2319	0.394	0.4268	0.315

Nano BEIR

• Dataset: NanoBEIR_mean
• Evaluated with NanoBEIREvaluator

Metric	Value
cosine_accuracy@1	0.3439
cosine_accuracy@3	0.535
cosine_accuracy@5	0.6013
cosine_accuracy@10	0.7122
cosine_precision@1	0.3439
cosine_precision@3	0.2345
cosine_precision@5	0.1843
cosine_precision@10	0.1341
cosine_recall@1	0.1858
cosine_recall@3	0.3258
cosine_recall@5	0.3834
cosine_recall@10	0.4845
cosine_ndcg@10	0.4081
cosine_mrr@10	0.4573
cosine_map@100	0.3386

Training Details

Training Dataset

gooaq-hard-negatives

• Dataset: gooaq-hard-negatives at 87594a1
• Size: 50,000 training samples
• Columns: question, answer, and negative
• Approximate statistics based on the first 1000 samples:

  	question	answer	negative
  type	string	string	string
  details	min: 8 tokens mean: 11.53 tokens max: 28 tokens	min: 14 tokens mean: 59.79 tokens max: 150 tokens	min: 15 tokens mean: 58.76 tokens max: 143 tokens

• Samples:

  question	answer	negative
  what is the difference between calories from fat and total fat?	Fat has more than twice as many calories per gram as carbohydrates and proteins. A gram of fat has about 9 calories, while a gram of carbohydrate or protein has about 4 calories. In other words, you could eat twice as much carbohydrates or proteins as fat for the same amount of calories.	Fat has more than twice as many calories per gram as carbohydrates and proteins. A gram of fat has about 9 calories, while a gram of carbohydrate or protein has about 4 calories. In other words, you could eat twice as much carbohydrates or proteins as fat for the same amount of calories.
  what is the difference between return transcript and account transcript?	A tax return transcript usually meets the needs of lending institutions offering mortgages and student loans. ... Tax Account Transcript - shows basic data such as return type, marital status, adjusted gross income, taxable income and all payment types. It also shows changes made after you filed your original return.	Trial balance is not a financial statement whereas a balance sheet is a financial statement. Trial balance is solely used for internal purposes whereas a balance sheet is used for purposes other than internal i.e. external. In a trial balance, each and every account is divided into debit (dr.) and credit (cr.)
  how long does my dog need to fast before sedation?	Now, guidelines are aimed towards 6-8 hours before surgery. This pre-op fasting time is much more beneficial for your pets because you have enough food in there to neutralize the stomach acid, preventing it from coming up the esophagus that causes regurgitation under anesthetic.	Try not to let your pooch rapidly wolf down his/her food! Do not let the dog play or exercise (e.g. go for a walk) for at least two hours after having a meal. Ensure continuous fresh water is available to avoid your pet gulping down a large amount after eating.

• Loss: CachedGISTEmbedLoss with these parameters:

  {'guide': SentenceTransformer(
    (0): Transformer({'max_seq_length': 256, 'do_lower_case': False}) with Transformer model: BertModel 
    (1): Pooling({'word_embedding_dimension': 384, '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()
  ), 'temperature': 0.01}
  

Evaluation Dataset

gooaq-hard-negatives

• Dataset: gooaq-hard-negatives at 87594a1
• Size: 10,048,700 evaluation samples
• Columns: question, answer, and negative
• Approximate statistics based on the first 1000 samples:

  	question	answer	negative
  type	string	string	string
  details	min: 8 tokens mean: 11.61 tokens max: 21 tokens	min: 16 tokens mean: 58.16 tokens max: 131 tokens	min: 14 tokens mean: 57.98 tokens max: 157 tokens

• Samples:

  question	answer	negative
  how is height width and length written?	The Graphics' industry standard is width by height (width x height). Meaning that when you write your measurements, you write them from your point of view, beginning with the width.	The Graphics' industry standard is width by height (width x height). Meaning that when you write your measurements, you write them from your point of view, beginning with the width. That's important.
  what is the difference between pork shoulder and loin?	All the recipes I've found for pulled pork recommends a shoulder/butt. Shoulders take longer to cook than a loin, because they're tougher. Loins are lean, while shoulders have marbled fat inside.	They are extracted from the loin, which runs from the hip to the shoulder, and it has a small strip of meat called the tenderloin. Unlike other pork, this pork chop is cut from four major sections, which are the shoulder, also known as the blade chops, ribs chops, loin chops, and the last, which is the sirloin chops.
  is the yin yang symbol religious?	The ubiquitous yin-yang symbol holds its roots in Taoism/Daoism, a Chinese religion and philosophy. The yin, the dark swirl, is associated with shadows, femininity, and the trough of a wave; the yang, the light swirl, represents brightness, passion and growth.	Yin energy is in the calm colors around you, in the soft music, in the soothing sound of a water fountain, or the relaxing images of water. Yang (active energy) is the feng shui energy expressed in strong, vibrant sounds and colors, bright lights, upward moving energy, tall plants, etc.

• Loss: CachedGISTEmbedLoss with these parameters:

  {'guide': SentenceTransformer(
    (0): Transformer({'max_seq_length': 256, 'do_lower_case': False}) with Transformer model: BertModel 
    (1): Pooling({'word_embedding_dimension': 384, '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()
  ), 'temperature': 0.01}
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 2048
• per_device_eval_batch_size: 2048
• learning_rate: 2e-05
• num_train_epochs: 1
• warmup_ratio: 0.1
• seed: 12
• bf16: True

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: 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
• 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
• use_liger_kernel: False
• eval_use_gather_object: False
• average_tokens_across_devices: False
• prompts: None
• batch_sampler: batch_sampler
• multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	Validation Loss	NanoClimateFEVER_cosine_ndcg@10	NanoDBPedia_cosine_ndcg@10	NanoFEVER_cosine_ndcg@10	NanoFiQA2018_cosine_ndcg@10	NanoHotpotQA_cosine_ndcg@10	NanoMSMARCO_cosine_ndcg@10	NanoNFCorpus_cosine_ndcg@10	NanoNQ_cosine_ndcg@10	NanoQuoraRetrieval_cosine_ndcg@10	NanoSCIDOCS_cosine_ndcg@10	NanoArguAna_cosine_ndcg@10	NanoSciFact_cosine_ndcg@10	NanoTouche2020_cosine_ndcg@10	NanoBEIR_mean_cosine_ndcg@10
0.04	1	11.5143	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.2	5	9.4399	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.4	10	5.5951	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.6	15	3.7416	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
0.8	20	2.8021	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
1.0	25	2.2003	1.3332	0.2733	0.3883	0.5162	0.3408	0.3824	0.3693	0.1698	0.3369	0.8834	0.2912	0.4950	0.4563	0.4021	0.4081

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.104 kWh
• Carbon Emitted: 0.040 kg of CO2
• Hours Used: 0.273 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: 3.5.0.dev0
• Transformers: 4.49.0
• PyTorch: 2.6.0+cu124
• Accelerate: 1.5.1
• Datasets: 3.3.2
• Tokenizers: 0.21.0

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",
}