tomaarsen/reranker-ModernBERT-base-gooaq-bce-no-pos-weight

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.

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-no-pos-weight")
# Get scores for pairs of texts
pairs = [
    ['what is a default final judgment?', 'Default judgment is a binding judgment in favor of either party based on some failure to take action by the other party. Most often, it is a judgment in favor of a plaintiff when the defendant has not responded to a summons or has failed to appear before a court of law. The failure to take action is the default.'],
    ['what is a default final judgment?', "A default judgment is a judgment issued against a party that doesn't bother to defend itself at all. ... A summary judgment is a judgment issued against a party that doesn't have any evidence to support its claims. Summary judgment means: “You can't prove it; therefore you lose.”"],
    ['what is a default final judgment?', 'This judgment is seen as being mentioned in Hebrews 9:27, which states that "it is appointed unto men once to die, but after this the judgment".'],
    ['what is a default final judgment?', "If you don't file an answer or go to court, your landlord can ask the judge to find you in default. Then the judge may let your landlord show there is reason for you to be evicted. If the landlord does that, the judge can enter a default judgment against you."],
    ['what is a default final judgment?', 'What can High Court Enforcement Officers do to enforce judgment? HCEOs can take control of goods or possessions to the value of the unpaid judgment, and may also attempt to take goods to cover the costs of enforcement, court costs, and interest on the debt.'],
]
scores = model.predict(pairs)
print(scores.shape)
# (5,)

# Or rank different texts based on similarity to a single text
ranks = model.rank(
    'what is a default final judgment?',
    [
        'Default judgment is a binding judgment in favor of either party based on some failure to take action by the other party. Most often, it is a judgment in favor of a plaintiff when the defendant has not responded to a summons or has failed to appear before a court of law. The failure to take action is the default.',
        "A default judgment is a judgment issued against a party that doesn't bother to defend itself at all. ... A summary judgment is a judgment issued against a party that doesn't have any evidence to support its claims. Summary judgment means: “You can't prove it; therefore you lose.”",
        'This judgment is seen as being mentioned in Hebrews 9:27, which states that "it is appointed unto men once to die, but after this the judgment".',
        "If you don't file an answer or go to court, your landlord can ask the judge to find you in default. Then the judge may let your landlord show there is reason for you to be evicted. If the landlord does that, the judge can enter a default judgment against you.",
        'What can High Court Enforcement Officers do to enforce judgment? HCEOs can take control of goods or possessions to the value of the unpaid judgment, and may also attempt to take goods to cover the costs of enforcement, court costs, and interest on the debt.',
    ]
)
# [{'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	0.7323 (+0.2012)
mrr@10	0.7309 (+0.2069)
ndcg@10	0.7731 (+0.1818)

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	0.4464 (-0.0431)	0.3794 (+0.1184)	0.5135 (+0.0939)
mrr@10	0.4352 (-0.0423)	0.5704 (+0.0706)	0.5180 (+0.0913)
ndcg@10	0.5250 (-0.0154)	0.4269 (+0.1018)	0.5685 (+0.0679)

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	0.4464 (+0.0564)
mrr@10	0.5079 (+0.0399)
ndcg@10	0.5068 (+0.0514)

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: 45.16 characters max: 84 characters	min: 51 characters mean: 252.6 characters max: 361 characters	0: ~82.80% 1: ~17.20%

• Samples:

  question	answer	label
  what is a default final judgment?	Default judgment is a binding judgment in favor of either party based on some failure to take action by the other party. Most often, it is a judgment in favor of a plaintiff when the defendant has not responded to a summons or has failed to appear before a court of law. The failure to take action is the default.	1
  what is a default final judgment?	A default judgment is a judgment issued against a party that doesn't bother to defend itself at all. ... A summary judgment is a judgment issued against a party that doesn't have any evidence to support its claims. Summary judgment means: “You can't prove it; therefore you lose.”	0
  what is a default final judgment?	This judgment is seen as being mentioned in Hebrews 9:27, which states that "it is appointed unto men once to die, but after this the judgment".	0

• Loss: BinaryCrossEntropyLoss with these parameters:

  {
      "activation_fn": "torch.nn.modules.linear.Identity",
      "pos_weight": null
  }
  

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	-	0.1386 (-0.4527)	0.0206 (-0.5198)	0.2387 (-0.0863)	0.0515 (-0.4491)	0.1036 (-0.3517)
0.0001	1	1.0425	-	-	-	-	-
0.0221	200	0.5627	-	-	-	-	-
0.0443	400	0.4593	-	-	-	-	-
0.0664	600	0.3714	-	-	-	-	-
0.0885	800	0.2955	-	-	-	-	-
0.1106	1000	0.2829	0.7083 (+0.1171)	0.4992 (-0.0412)	0.3110 (-0.0141)	0.4795 (-0.0211)	0.4299 (-0.0255)
0.1328	1200	0.2696	-	-	-	-	-
0.1549	1400	0.2548	-	-	-	-	-
0.1770	1600	0.2485	-	-	-	-	-
0.1992	1800	0.2326	-	-	-	-	-
0.2213	2000	0.241	0.7461 (+0.1549)	0.5350 (-0.0054)	0.3701 (+0.0451)	0.5339 (+0.0332)	0.4797 (+0.0243)
0.2434	2200	0.2373	-	-	-	-	-
0.2655	2400	0.2356	-	-	-	-	-
0.2877	2600	0.2207	-	-	-	-	-
0.3098	2800	0.222	-	-	-	-	-
0.3319	3000	0.2258	0.7443 (+0.1531)	0.5554 (+0.0150)	0.3921 (+0.0671)	0.5368 (+0.0361)	0.4948 (+0.0394)
0.3541	3200	0.2182	-	-	-	-	-
0.3762	3400	0.215	-	-	-	-	-
0.3983	3600	0.2161	-	-	-	-	-
0.4204	3800	0.2202	-	-	-	-	-
0.4426	4000	0.2147	0.7542 (+0.1629)	0.5465 (+0.0061)	0.4047 (+0.0797)	0.5199 (+0.0193)	0.4904 (+0.0350)
0.4647	4200	0.2177	-	-	-	-	-
0.4868	4400	0.2129	-	-	-	-	-
0.5090	4600	0.2099	-	-	-	-	-
0.5311	4800	0.2105	-	-	-	-	-
0.5532	5000	0.2101	0.7644 (+0.1731)	0.5448 (+0.0044)	0.4157 (+0.0907)	0.5746 (+0.0739)	0.5117 (+0.0563)
0.5753	5200	0.2034	-	-	-	-	-
0.5975	5400	0.2047	-	-	-	-	-
0.6196	5600	0.2043	-	-	-	-	-
0.6417	5800	0.2029	-	-	-	-	-
0.6639	6000	0.2021	0.7699 (+0.1786)	0.5250 (-0.0154)	0.4264 (+0.1013)	0.5491 (+0.0484)	0.5002 (+0.0448)
0.6860	6200	0.2048	-	-	-	-	-
0.7081	6400	0.2033	-	-	-	-	-
0.7303	6600	0.2017	-	-	-	-	-
0.7524	6800	0.1976	-	-	-	-	-
0.7745	7000	0.1989	0.7722 (+0.1810)	0.5732 (+0.0328)	0.4206 (+0.0956)	0.6013 (+0.1007)	0.5317 (+0.0763)
0.7966	7200	0.1925	-	-	-	-	-
0.8188	7400	0.1917	-	-	-	-	-
0.8409	7600	0.2002	-	-	-	-	-
0.8630	7800	0.1913	-	-	-	-	-
0.8852	8000	0.191	0.7707 (+0.1794)	0.5412 (+0.0007)	0.4332 (+0.1082)	0.5508 (+0.0502)	0.5084 (+0.0530)
0.9073	8200	0.1929	-	-	-	-	-
0.9294	8400	0.1989	-	-	-	-	-
0.9515	8600	0.1889	-	-	-	-	-
0.9737	8800	0.1874	-	-	-	-	-
0.9958	9000	0.1863	0.7731 (+0.1818)	0.5250 (-0.0154)	0.4269 (+0.1018)	0.5685 (+0.0679)	0.5068 (+0.0514)
-1	-1	-	0.7731 (+0.1818)	0.5250 (-0.0154)	0.4269 (+0.1018)	0.5685 (+0.0679)	0.5068 (+0.0514)

• 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",
}