tomaarsen/reranker-msmarco-v1.1-MiniLM-L12-H384-uncased-plistmle-softmax

CrossEncoder based on microsoft/MiniLM-L12-H384-uncased

This is a Cross Encoder model finetuned from microsoft/MiniLM-L12-H384-uncased on the ms_marco dataset using the sentence-transformers library. It computes scores for pairs of texts, which can be used for text reranking and semantic search.

Model Details

Model Description

• Model Type: Cross Encoder
• Base model: microsoft/MiniLM-L12-H384-uncased
• Maximum Sequence Length: 512 tokens
• Number of Output Labels: 1 label
• Training Dataset:
  • ms_marco
• Language: en

Model Sources

• Documentation: Sentence Transformers Documentation
• Documentation: Cross Encoder Documentation
• Repository: Sentence Transformers on GitHub
• Hugging Face: Cross Encoders on Hugging Face

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 CrossEncoder

# Download from the 🤗 Hub
model = CrossEncoder("tomaarsen/reranker-msmarco-v1.1-MiniLM-L12-H384-uncased-plistmle-softmax")
# Get scores for pairs of texts
pairs = [
    ['How many calories in an egg', 'There are on average between 55 and 80 calories in an egg depending on its size.'],
    ['How many calories in an egg', 'Egg whites are very low in calories, have no fat, no cholesterol, and are loaded with protein.'],
    ['How many calories in an egg', 'Most of the calories in an egg come from the yellow yolk in the center.'],
]
scores = model.predict(pairs)
print(scores.shape)
# (3,)

# Or rank different texts based on similarity to a single text
ranks = model.rank(
    'How many calories in an egg',
    [
        'There are on average between 55 and 80 calories in an egg depending on its size.',
        'Egg whites are very low in calories, have no fat, no cholesterol, and are loaded with protein.',
        'Most of the calories in an egg come from the yellow yolk in the center.',
    ]
)
# [{'corpus_id': ..., 'score': ...}, {'corpus_id': ..., 'score': ...}, ...]

Evaluation

Metrics

Cross Encoder Reranking

• Datasets: NanoMSMARCO_R100, NanoNFCorpus_R100 and NanoNQ_R100
• Evaluated with CrossEncoderRerankingEvaluator with these parameters:

  {
      "at_k": 10,
      "always_rerank_positives": true
  }
  

Metric	NanoMSMARCO_R100	NanoNFCorpus_R100	NanoNQ_R100
map	0.5177 (+0.0281)	0.3178 (+0.0568)	0.6038 (+0.1842)
mrr@10	0.5097 (+0.0322)	0.5947 (+0.0949)	0.6118 (+0.1851)
ndcg@10	0.5896 (+0.0492)	0.3558 (+0.0308)	0.6704 (+0.1697)

Cross Encoder Nano BEIR

• Dataset: NanoBEIR_R100_mean
• Evaluated with CrossEncoderNanoBEIREvaluator with these parameters:

  {
      "dataset_names": [
          "msmarco",
          "nfcorpus",
          "nq"
      ],
      "rerank_k": 100,
      "at_k": 10,
      "always_rerank_positives": true
  }
  

Metric	Value
map	0.4797 (+0.0897)
mrr@10	0.5721 (+0.1040)
ndcg@10	0.5386 (+0.0832)

Training Details

Training Dataset

ms_marco

• Dataset: ms_marco at a47ee7a
• Size: 78,704 training samples
• Columns: query, docs, and labels
• Approximate statistics based on the first 1000 samples:

  	query	docs	labels
  type	string	list	list
  details	min: 10 characters mean: 34.08 characters max: 109 characters	min: 3 elements mean: 6.50 elements max: 10 elements	min: 3 elements mean: 6.50 elements max: 10 elements

• Samples:

  query	docs	labels
  what is ha	['1 ha. interjection \ˈhae\. ˈhä. —tweet used especially to express surprise or a feeling of pleasure that you have when you do something or find out about. something Full Definition of HA. —used especially to express surprise, joy, or triumph. See ha defined for English-language learners.', 'Ha-ha protecting the lawn at Heaton Hall, Manchester. Note how the wall disappears as it curves away to the right of the picture. A ha-ha is a recessed landscape design element that creates a vertical barrier whilst preserving an uninterrupted view of the landscape beyond. The design includes a turfed incline which slopes downward to a sharply vertical face, typically a masonry retaining wall. Ha-has are used in landscape design to prevent access to a garden, for example by grazing livestock, without obstructing views. In security design, the element is used to deter vehicular access to a site while minimizing visual obstruction', 'The doctors said they are afraid of HA-MRSA. Why is it so scary?...	[1, 0, 0, 0, 0, ...]
  salary for kinesiology degree	['Coaches. Coaching is another common career option for those holding a degree in kinesiology. Athletic coaches made an average annual salary of $35,950, as of May 2010, according to the bureau, while the middle 50 percent made between $18,800 and $43,930 annually.', "Athletic Trainers. One common career option for those holding a bachelor's degree in kinesiology is a career in the athletic training field. Athletic trainers made an average annual salary of $44,030, as of May 2010, according to the U.S. Bureau of Labor Statistics.", 'Kinesiologist salary. A newly graduated kinesiologist may receive a yearly minimum pay of around $20,000. This average will increase significantly after a year. As a matter of fact a kinesiologist can receive up to around $80,000 yearly. The median average salary of all kinesiologists is around $46,000 per annum.', 'Salary of starting kinesiologist in Georgia ranges from $29,592 – $44,387. The average kinesiology wage ranges from $35,933 – $53,899 and the t...	[1, 0, 0, 0, 0, ...]
  average temperatures owensboro ky	['2.8. The highest average temperature in Owensboro is August at 79.5 degrees. The coldest average temperature in Owensboro is February at 37 degrees. The most monthly precipitation in Owensboro occurs in September with 4 inches. The Owensboro weather information is based on the average of the previous 3-7 years of data. Loading...', 'Most / Least Educated Cities in KY. The average temperature of Owensboro is 56.97°F, which is higher than the Kentucky average temperature of 55.62°F and is higher than the national average temperature of 54.45°F.', 'Owensboro, Kentucky, gets 45 inches of rain per year. The US average is 37. Snowfall is 10 inches. The average US city gets 25 inches of snow per year. The number of days with any measurable precipitation is 103. On average, there are 202 sunny days per year in Owensboro, Kentucky. The July high is around 90 degrees. The January low is 25.']	[1, 0, 0]

• Loss: PListMLELoss with these parameters:

  {
      "lambda_weight": "sentence_transformers.cross_encoder.losses.PListMLELoss.PListMLELambdaWeight",
      "activation_fct": "torch.nn.modules.linear.Identity",
      "mini_batch_size": 16,
      "respect_input_order": true
  }
  

Evaluation Dataset

ms_marco

• Dataset: ms_marco at a47ee7a
• Size: 1,000 evaluation samples
• Columns: query, docs, and labels
• Approximate statistics based on the first 1000 samples:

  	query	docs	labels
  type	string	list	list
  details	min: 8 characters mean: 33.76 characters max: 90 characters	min: 2 elements mean: 6.00 elements max: 10 elements	min: 2 elements mean: 6.00 elements max: 10 elements

• Samples:

  query	docs	labels
  meaning of regional value content	['Regional Value Content Some specific rules of origin require that a good have a minimum regional value content, meaning that a certain percentage of the value of the goods must be from North America. There are two formulas for calculating the regional value content. TV. Where RVC is the regional value content, expressed as a percentage; TV is the transaction value of the good adjusted to an FOB basis; and VNM is the value of non-originating material used by the producer in the production of the good.', 'Regional Value Content: For goods subject to a regional value content rule of origin, please refer to Regional Value content (Explanatory Material and Article 402), for an explanation of regional value content rules of origin. In the two above situations, no tariff shift is possible because of how the goods were classified. Goods in this situation may obtain NAFTA tariff preference if they have 50 or 60 percent North American value content, depending on method used. Refer to Regional ...	[1, 0, 0, 0, 0, ...]
  how long does it take to get an irs refund	['Essentially, they’re saying most people will get their refunds in less than 21 days, but there won’t be any schedules they will follow this year. The IRS provides these tips to tax preparers to get the fastest tax refund: 1 File an accurate tax return. 2 Efile your tax return. The IRS eliminated the refund schedules that were used in prior years for both direct deposit and mailed refunds. The guideline the IRS is using in 2015 for filing your 2014 tax return is: Don’t count on getting your refund by a certain date to make major purchases or pay other financial obligations. Even though the IRS issues most refunds in less than 21 days, it’s possible your tax return may require additional review and take longer', 'If you file a paper return, the IRS says you should allow about six weeks to receive your refund. If you file Form 8379, Injured Spouse Allocation, it could take up to 14 weeks to process your tax return. ', 'Tax refunds are normally issued within approximately 21 days if yo...	[1, 0, 0, 0, 0, ...]
  calculate margin on cost to retail	['The formula for calculating retail margin is the sales price of an item minus COGS, divided by the sales price, multiplied by 100. If you sell an item at $20 and paid $10 to acquire it and sell it, your retail margin is $10 divided by $20, or 50 percent. You subtract your COGS from your desired sales price in the same way, but then divide that amount by your COGS. If your target price is $20 and your COGS are $10, your markup is $10 divided by $10, or 100 percent. Thus, to achieve a 50 percent margin on an item that costs you $10, you need a 100 percent markup.', 'This gives you a 70 percent cost-to-retail ratio. For the year, you book $336,000 in net sales. Your cost of goods sold under the retail method is 70 percent of net sales, or $235,200. Your gross profit is $100,800, which is 30 percent of net sales. You can shortcut the calculation of gross profit margin by subtracting your cost-to-retail percentage of 70 percent from 100 percent. While easy enough to calculate, the gross m...	[1, 1, 0, 0, 0, ...]

• Loss: PListMLELoss with these parameters:

  {
      "lambda_weight": "sentence_transformers.cross_encoder.losses.PListMLELoss.PListMLELambdaWeight",
      "activation_fct": "torch.nn.modules.linear.Identity",
      "mini_batch_size": 16,
      "respect_input_order": true
  }
  

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
• load_best_model_at_end: 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: 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: True
• 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	NanoMSMARCO_R100_ndcg@10	NanoNFCorpus_R100_ndcg@10	NanoNQ_R100_ndcg@10	NanoBEIR_R100_mean_ndcg@10
-1	-1	-	-	0.0785 (-0.4620)	0.2441 (-0.0809)	0.0473 (-0.4533)	0.1233 (-0.3321)
0.0002	1	2.2881	-	-	-	-	-
0.0508	250	2.2177	-	-	-	-	-
0.1016	500	2.0172	1.8991	0.3929 (-0.1475)	0.2715 (-0.0535)	0.3910 (-0.1096)	0.3518 (-0.1036)
0.1525	750	1.8769	-	-	-	-	-
0.2033	1000	1.8053	1.7602	0.5109 (-0.0296)	0.3391 (+0.0140)	0.6059 (+0.1052)	0.4853 (+0.0299)
0.2541	1250	1.7566	-	-	-	-	-
0.3049	1500	1.7298	1.7028	0.5152 (-0.0253)	0.3334 (+0.0084)	0.6332 (+0.1326)	0.4939 (+0.0385)
0.3558	1750	1.7261	-	-	-	-	-
0.4066	2000	1.6907	1.7166	0.5441 (+0.0037)	0.3571 (+0.0321)	0.6231 (+0.1224)	0.5081 (+0.0527)
0.4574	2250	1.6579	-	-	-	-	-
0.5082	2500	1.6599	1.6482	0.5434 (+0.0030)	0.3284 (+0.0034)	0.6008 (+0.1001)	0.4909 (+0.0355)
0.5591	2750	1.6467	-	-	-	-	-
0.6099	3000	1.6489	1.6266	0.5541 (+0.0137)	0.3514 (+0.0264)	0.6354 (+0.1347)	0.5136 (+0.0583)
0.6607	3250	1.6538	-	-	-	-	-
0.7115	3500	1.6246	1.6389	0.5896 (+0.0492)	0.3558 (+0.0308)	0.6704 (+0.1697)	0.5386 (+0.0832)
0.7624	3750	1.6164	-	-	-	-	-
0.8132	4000	1.6139	1.6174	0.5639 (+0.0235)	0.3358 (+0.0108)	0.5988 (+0.0981)	0.4995 (+0.0441)
0.8640	4250	1.6283	-	-	-	-	-
0.9148	4500	1.6265	1.5999	0.5737 (+0.0333)	0.3412 (+0.0161)	0.6161 (+0.1154)	0.5103 (+0.0550)
0.9656	4750	1.5755	-	-	-	-	-
-1	-1	-	-	0.5896 (+0.0492)	0.3558 (+0.0308)	0.6704 (+0.1697)	0.5386 (+0.0832)

• The bold row denotes the saved checkpoint.

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 0.236 kWh
• Carbon Emitted: 0.092 kg of CO2
• Hours Used: 0.882 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",
}

PListMLELoss

@inproceedings{lan2014position,
  title={Position-Aware ListMLE: A Sequential Learning Process for Ranking.},
  author={Lan, Yanyan and Zhu, Yadong and Guo, Jiafeng and Niu, Shuzi and Cheng, Xueqi},
  booktitle={UAI},
  volume={14},
  pages={449--458},
  year={2014}
}