tomaarsen/reranker-ModernBERT-base-gooaq-bce

ModernBERT-base trained on GooAQ

This is a Cross Encoder model finetuned from answerdotai/ModernBERT-base using the sentence-transformers library. It computes scores for pairs of texts, which can be used for text reranking and semantic search.

See training_gooaq_bce.py for the training script. This script is also described in the Cross Encoder > Training Overview documentation and the Training and Finetuning Reranker Models with Sentence Transformers v4 blogpost.

Model Details

Model Description

• Model Type: Cross Encoder
• Base model: answerdotai/ModernBERT-base
• Maximum Sequence Length: 8192 tokens
• Number of Output Labels: 1 label
• Language: en
• License: apache-2.0

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-ModernBERT-base-gooaq-bce")
# Get scores for pairs of texts
pairs = [
    ['why are rye chips so good?', "It makes them taste that much better! The rye chips are tasty because they stand out--they're the saltiest thing in the bag. It's not because rye bread is inherently awesome. ... You could just buy a bag of rye chips."],
    ['why are rye chips so good?', 'There are no substantial technical, nutritional or performance issues associated with rye that would limit its use for pets. Rye is a fairly common ingredient in human foods and beverages. The most prevalent occurrence is in crackers and breads.'],
    ['why are rye chips so good?', 'Bread made wholly from rye flour is made in Germany and called pumpernickel. Rye is unique among grains for having a high level of fibre in its endosperm – not just in its bran. As such, the glycemic index (GI) of rye products is generally lower than products made from wheat and most other grains.'],
    ['why are rye chips so good?', 'KFC Chips – The salt mix on the seasoned chips and the actual chips do not contain any animal products. Our supplier/s of chips and seasoning have confirmed they are suitable for vegans.'],
    ['why are rye chips so good?', 'A study in the American Journal of Clinical Nutrition found that eating rye leads to better blood-sugar control compared to wheat. Rye bread is packed with magnesium, which helps control blood pressure and optimize heart health. Its high levels of soluble fibre can also reduce cholesterol.'],
]
scores = model.predict(pairs)
print(scores.shape)
# (5,)

# Or rank different texts based on similarity to a single text
ranks = model.rank(
    'why are rye chips so good?',
    [
        "It makes them taste that much better! The rye chips are tasty because they stand out--they're the saltiest thing in the bag. It's not because rye bread is inherently awesome. ... You could just buy a bag of rye chips.",
        'There are no substantial technical, nutritional or performance issues associated with rye that would limit its use for pets. Rye is a fairly common ingredient in human foods and beverages. The most prevalent occurrence is in crackers and breads.',
        'Bread made wholly from rye flour is made in Germany and called pumpernickel. Rye is unique among grains for having a high level of fibre in its endosperm – not just in its bran. As such, the glycemic index (GI) of rye products is generally lower than products made from wheat and most other grains.',
        'KFC Chips – The salt mix on the seasoned chips and the actual chips do not contain any animal products. Our supplier/s of chips and seasoning have confirmed they are suitable for vegans.',
        'A study in the American Journal of Clinical Nutrition found that eating rye leads to better blood-sugar control compared to wheat. Rye bread is packed with magnesium, which helps control blood pressure and optimize heart health. Its high levels of soluble fibre can also reduce cholesterol.',
    ]
)
# [{'corpus_id': ..., 'score': ...}, {'corpus_id': ..., 'score': ...}, ...]

Evaluation

Metrics

Cross Encoder Reranking

• Dataset: gooaq-dev
• Evaluated with CrossEncoderRerankingEvaluator with these parameters:

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

Metric	Value
map	0.7308 (+0.1997)
mrr@10	0.7292 (+0.2052)
ndcg@10	0.7713 (+0.1801)

Cross Encoder Reranking

• Dataset: gooaq-dev
• Evaluated with CrossEncoderRerankingEvaluator with these parameters:

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

Metric	Value
map	0.7908 (+0.2597)
mrr@10	0.7890 (+0.2650)
ndcg@10	0.8351 (+0.2439)

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	0.4579 (-0.0317)	0.3414 (+0.0804)	0.3932 (-0.0264)
mrr@10	0.4479 (-0.0296)	0.5340 (+0.0342)	0.3918 (-0.0349)
ndcg@10	0.5275 (-0.0130)	0.3821 (+0.0571)	0.4630 (-0.0377)

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	0.3975 (+0.0074)
mrr@10	0.4579 (-0.0101)
ndcg@10	0.4575 (+0.0022)

Training Details

Training Dataset

Unnamed Dataset

• Size: 578,402 training samples
• Columns: question, answer, and label
• Approximate statistics based on the first 1000 samples:

  	question	answer	label
  type	string	string	int
  details	min: 19 characters mean: 45.14 characters max: 85 characters	min: 65 characters mean: 254.8 characters max: 379 characters	0: ~82.90% 1: ~17.10%

• Samples:

  question	answer	label
  why are rye chips so good?	It makes them taste that much better! The rye chips are tasty because they stand out--they're the saltiest thing in the bag. It's not because rye bread is inherently awesome. ... You could just buy a bag of rye chips.	1
  why are rye chips so good?	There are no substantial technical, nutritional or performance issues associated with rye that would limit its use for pets. Rye is a fairly common ingredient in human foods and beverages. The most prevalent occurrence is in crackers and breads.	0
  why are rye chips so good?	Bread made wholly from rye flour is made in Germany and called pumpernickel. Rye is unique among grains for having a high level of fibre in its endosperm – not just in its bran. As such, the glycemic index (GI) of rye products is generally lower than products made from wheat and most other grains.	0

• Loss: BinaryCrossEntropyLoss with these parameters:

  {
      "activation_fct": "torch.nn.modules.linear.Identity",
      "pos_weight": 5
  }
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 64
• per_device_eval_batch_size: 64
• learning_rate: 2e-05
• num_train_epochs: 1
• warmup_ratio: 0.1
• seed: 12
• bf16: True
• dataloader_num_workers: 4
• 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: 64
• per_device_eval_batch_size: 64
• 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: 4
• 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	gooaq-dev_ndcg@10	NanoMSMARCO_R100_ndcg@10	NanoNFCorpus_R100_ndcg@10	NanoNQ_R100_ndcg@10	NanoBEIR_R100_mean_ndcg@10
-1	-1	-	0.1288 (-0.4624)	0.0149 (-0.5255)	0.2278 (-0.0972)	0.0229 (-0.4777)	0.0885 (-0.3668)
0.0001	1	1.0435	-	-	-	-	-
0.0221	200	1.1924	-	-	-	-	-
0.0443	400	1.1531	-	-	-	-	-
0.0664	600	0.9371	-	-	-	-	-
0.0885	800	0.6993	-	-	-	-	-
0.1106	1000	0.669	0.7042 (+0.1130)	0.4353 (-0.1051)	0.3289 (+0.0039)	0.4250 (-0.0757)	0.3964 (-0.0590)
0.1328	1200	0.6257	-	-	-	-	-
0.1549	1400	0.6283	-	-	-	-	-
0.1770	1600	0.6014	-	-	-	-	-
0.1992	1800	0.5888	-	-	-	-	-
0.2213	2000	0.5493	0.7425 (+0.1513)	0.4947 (-0.0457)	0.3568 (+0.0318)	0.4634 (-0.0373)	0.4383 (-0.0171)
0.2434	2200	0.5479	-	-	-	-	-
0.2655	2400	0.5329	-	-	-	-	-
0.2877	2600	0.5208	-	-	-	-	-
0.3098	2800	0.5259	-	-	-	-	-
0.3319	3000	0.5221	0.7479 (+0.1567)	0.5146 (-0.0258)	0.3710 (+0.0460)	0.4846 (-0.0160)	0.4568 (+0.0014)
0.3541	3200	0.4977	-	-	-	-	-
0.3762	3400	0.4965	-	-	-	-	-
0.3983	3600	0.4985	-	-	-	-	-
0.4204	3800	0.4907	-	-	-	-	-
0.4426	4000	0.5058	0.7624 (+0.1712)	0.5166 (-0.0238)	0.3665 (+0.0415)	0.4868 (-0.0138)	0.4567 (+0.0013)
0.4647	4200	0.4885	-	-	-	-	-
0.4868	4400	0.495	-	-	-	-	-
0.5090	4600	0.4839	-	-	-	-	-
0.5311	4800	0.4983	-	-	-	-	-
0.5532	5000	0.4778	0.7603 (+0.1691)	0.5110 (-0.0294)	0.3540 (+0.0290)	0.4809 (-0.0197)	0.4487 (-0.0067)
0.5753	5200	0.4726	-	-	-	-	-
0.5975	5400	0.477	-	-	-	-	-
0.6196	5600	0.4613	-	-	-	-	-
0.6417	5800	0.4492	-	-	-	-	-
0.6639	6000	0.4506	0.7643 (+0.1731)	0.5275 (-0.0129)	0.3639 (+0.0389)	0.4913 (-0.0094)	0.4609 (+0.0055)
0.6860	6200	0.4618	-	-	-	-	-
0.7081	6400	0.463	-	-	-	-	-
0.7303	6600	0.4585	-	-	-	-	-
0.7524	6800	0.4612	-	-	-	-	-
0.7745	7000	0.4621	0.7649 (+0.1736)	0.5105 (-0.0299)	0.3688 (+0.0437)	0.4552 (-0.0454)	0.4448 (-0.0105)
0.7966	7200	0.4536	-	-	-	-	-
0.8188	7400	0.4515	-	-	-	-	-
0.8409	7600	0.4396	-	-	-	-	-
0.8630	7800	0.4542	-	-	-	-	-
0.8852	8000	0.4332	0.7669 (+0.1757)	0.5247 (-0.0157)	0.3794 (+0.0544)	0.4370 (-0.0637)	0.4470 (-0.0083)
0.9073	8200	0.447	-	-	-	-	-
0.9294	8400	0.4335	-	-	-	-	-
0.9515	8600	0.4179	-	-	-	-	-
0.9737	8800	0.4459	-	-	-	-	-
0.9958	9000	0.4196	0.7713 (+0.1801)	0.5275 (-0.0130)	0.3821 (+0.0571)	0.4630 (-0.0377)	0.4575 (+0.0022)
-1	-1	-	0.7713 (+0.1801)	0.5275 (-0.0130)	0.3821 (+0.0571)	0.4630 (-0.0377)	0.4575 (+0.0022)

• The bold row denotes the saved checkpoint.

Framework Versions

• Python: 3.11.10
• Sentence Transformers: 3.5.0.dev0
• Transformers: 4.49.0
• PyTorch: 2.5.1+cu124
• Accelerate: 1.5.2
• Datasets: 2.21.0
• 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",
}