Add new CrossEncoder model

c1cdb06 verified 3 days ago

28.9 kB

language:
  - en
license: apache-2.0
tags:
  - sentence-transformers
  - cross-encoder
  - reranker
  - generated_from_trainer
  - dataset_size:95939
  - loss:LambdaLoss
base_model: answerdotai/ModernBERT-base
pipeline_tag: text-ranking
library_name: sentence-transformers
metrics:
  - map
  - mrr@10
  - ndcg@10
model-index:
  - name: ModernBERT-base trained on GooAQ
    results:
      - task:
          type: cross-encoder-reranking
          name: Cross Encoder Reranking
        dataset:
          name: gooaq dev
          type: gooaq-dev
        metrics:
          - type: map
            value: 0.7235
            name: Map
          - type: mrr@10
            value: 0.722
            name: Mrr@10
          - type: ndcg@10
            value: 0.7664
            name: Ndcg@10
      - task:
          type: cross-encoder-reranking
          name: Cross Encoder Reranking
        dataset:
          name: NanoMSMARCO R100
          type: NanoMSMARCO_R100
        metrics:
          - type: map
            value: 0.44
            name: Map
          - type: mrr@10
            value: 0.4276
            name: Mrr@10
          - type: ndcg@10
            value: 0.5023
            name: Ndcg@10
      - task:
          type: cross-encoder-reranking
          name: Cross Encoder Reranking
        dataset:
          name: NanoNFCorpus R100
          type: NanoNFCorpus_R100
        metrics:
          - type: map
            value: 0.3312
            name: Map
          - type: mrr@10
            value: 0.5754
            name: Mrr@10
          - type: ndcg@10
            value: 0.3743
            name: Ndcg@10
      - task:
          type: cross-encoder-reranking
          name: Cross Encoder Reranking
        dataset:
          name: NanoNQ R100
          type: NanoNQ_R100
        metrics:
          - type: map
            value: 0.2867
            name: Map
          - type: mrr@10
            value: 0.281
            name: Mrr@10
          - type: ndcg@10
            value: 0.3538
            name: Ndcg@10
      - task:
          type: cross-encoder-nano-beir
          name: Cross Encoder Nano BEIR
        dataset:
          name: NanoBEIR R100 mean
          type: NanoBEIR_R100_mean
        metrics:
          - type: map
            value: 0.3526
            name: Map
          - type: mrr@10
            value: 0.428
            name: Mrr@10
          - type: ndcg@10
            value: 0.4101
            name: Ndcg@10

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
Supported Modality: Text
Language: en
License: apach-2.0

Model Sources

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

Full Model Architecture

CrossEncoder(
  (0): Transformer({'transformer_task': 'sequence-classification', 'modality_config': {'text': {'method': 'forward', 'method_output_name': 'logits'}}, 'module_output_name': 'scores', 'architecture': 'ModernBertForSequenceClassification'})
)

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("cross_encoder_model_id")
# Get scores for pairs of inputs
pairs = [
    ['what are the 5 characteristics of a star?', 'Key Concept: Characteristics used to classify stars include color, temperature, size, composition, and brightness.'],
    ['what are the 5 characteristics of a star?', 'The biggest difference was compression. The most recent Z-STAR has a compression of 90, while the Z-STAR XV comes in at 102. The original Q-STAR Tour had a compression of 75 (this new version, like Gen 2, is 72), and with its performance characteristics, Srixon could have called it Z-STAR Soft.'],
    ['what are the 5 characteristics of a star?', "While 4-star hotels provide quality services to its patrons, a 5-star luxury accommodation will have impeccable service delivery. You'll often end up paying extra for the amenities like room service and concierge services in a 4-star hotel. But a 5-star hotel will always offer personalised services."],
    ['what are the 5 characteristics of a star?', 'However, the brightness of a star depends on its composition and how far it is from the planet. Astronomers define star brightness in terms of apparent magnitude — how bright the star appears from Earth — and absolute magnitude — how bright the star appears at a standard distance of 32.6 light-years, or 10 parsecs.'],
    ['what are the 5 characteristics of a star?', 'A "falling star" or a "shooting star" has nothing at all to do with a star! These amazing streaks of light you can sometimes see in the night sky are caused by tiny bits of dust and rock called meteoroids falling into the Earth\'s atmosphere and burning up. ... Meteors are commonly called falling stars or shooting stars.'],
]
scores = model.predict(pairs)
print(scores)
# [9.4520e-01 1.7401e-04 4.5739e-03 2.9579e-03 1.3631e-05]

# Or rank different texts based on similarity to a single text
ranks = model.rank(
    'what are the 5 characteristics of a star?',
    [
        'Key Concept: Characteristics used to classify stars include color, temperature, size, composition, and brightness.',
        'The biggest difference was compression. The most recent Z-STAR has a compression of 90, while the Z-STAR XV comes in at 102. The original Q-STAR Tour had a compression of 75 (this new version, like Gen 2, is 72), and with its performance characteristics, Srixon could have called it Z-STAR Soft.',
        "While 4-star hotels provide quality services to its patrons, a 5-star luxury accommodation will have impeccable service delivery. You'll often end up paying extra for the amenities like room service and concierge services in a 4-star hotel. But a 5-star hotel will always offer personalised services.",
        'However, the brightness of a star depends on its composition and how far it is from the planet. Astronomers define star brightness in terms of apparent magnitude — how bright the star appears from Earth — and absolute magnitude — how bright the star appears at a standard distance of 32.6 light-years, or 10 parsecs.',
        'A "falling star" or a "shooting star" has nothing at all to do with a star! These amazing streaks of light you can sometimes see in the night sky are caused by tiny bits of dust and rock called meteoroids falling into the Earth\'s atmosphere and burning up. ... Meteors are commonly called falling stars or shooting stars.',
    ]
)
# [{'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	0.7235 (+0.1924)
mrr@10	0.7220 (+0.1981)
ndcg@10	0.7664 (+0.1751)

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	0.4400 (-0.0496)	0.3312 (+0.0702)	0.2867 (-0.1329)
mrr@10	0.4276 (-0.0499)	0.5754 (+0.0756)	0.2810 (-0.1457)
ndcg@10	0.5023 (-0.0381)	0.3743 (+0.0492)	0.3538 (-0.1468)

Cross Encoder Nano BEIR

Dataset: NanoBEIR_R100_mean

Evaluated with CrossEncoderNanoBEIREvaluator with these parameters:

{
    "dataset_names": [
        "msmarco",
        "nfcorpus",
        "nq"
    ],
    "dataset_id": "sentence-transformers/NanoBEIR-en",
    "rerank_k": 100,
    "at_k": 10,
    "always_rerank_positives": true
}

Metric	Value
map	0.3526 (-0.0374)
mrr@10	0.4280 (-0.0400)
ndcg@10	0.4101 (-0.0452)

Training Details

Training Dataset

Unnamed Dataset

Size: 95,939 training samples
Columns: question, answer, and labels
Approximate statistics based on the first 100 samples:
question answer labels
type string list list

modality text
details
min: 8 tokens
mean: 11.69 tokens
max: 19 tokens

size: 6 elements

size: 6 elements

	question	answer	labels
type	string	list	list
modality	text
details	min: 8 tokens mean: 11.69 tokens max: 19 tokens	size: 6 elements	size: 6 elements

Samples:

question	answer	labels
`what are the 5 characteristics of a star?`	['Key Concept: Characteristics used to classify stars include color, temperature, size, composition, and brightness.', 'The biggest difference was compression. The most recent Z-STAR has a compression of 90, while the Z-STAR XV comes in at 102. The original Q-STAR Tour had a compression of 75 (this new version, like Gen 2, is 72), and with its performance characteristics, Srixon could have called it Z-STAR Soft.', "While 4-star hotels provide quality services to its patrons, a 5-star luxury accommodation will have impeccable service delivery. You'll often end up paying extra for the amenities like room service and concierge services in a 4-star hotel. But a 5-star hotel will always offer personalised services.", 'However, the brightness of a star depends on its composition and how far it is from the planet. Astronomers define star brightness in terms of apparent magnitude — how bright the star appears from Earth — and absolute magnitude — how bright the star appears at a standard distance of 32.6 light-years, or 10 parsecs.', 'A "falling star" or a "shooting star" has nothing at all to do with a star! These amazing streaks of light you can sometimes see in the night sky are caused by tiny bits of dust and rock called meteoroids falling into the Earth's atmosphere and burning up. ... Meteors are commonly called falling stars or shooting stars.', ...]	`[1, 0, 0, 0, 0, ...]`
`are copic markers alcohol ink?`	['Copic Ink is alcohol-based and flammable. Keep away from direct sunlight and extreme temperatures.', 'Copic Sketch markers are alcohol-based, refillable dye markers with replaceable nibs for full versatility. ... Great for comic drawing, scrapbooking, fine art, illustration, product design, architecture, and more, this marker is an essential for artists, designers and crafters. Details: Available in multiple colors.', 'Alcohol based markers differ from water based markers in that the color (dye or pigment) is suspended in an alcohol or another fast evaporating solvent, rather than water or glycerin. This means that alcohol based markers are not water soluble, but may be alcohol soluble.', 'Gel ink pens are primarily made to be used on paper substrates and not on plastics, glass, or fabrics. While some gel ink versions can adhere to other surfaces, this greatly affects the overall "performance" of the ink. (i.e. the gel inks might be easily washed out or scratched off the surface.)', 'Permanent markers are made from pigment and not dyes. Dyes are water soluble but pigment are not soluble in water. If you clean the markers with a no polar solvent like alcohol or acetone the markers are not truly permanent on certain materials. ... The purpose is to determine if permanent marker are truly permanent.', ...]	`[1, 0, 0, 0, 0, ...]`
`what is the difference between appellate term and appellate division?`	['Appellate terms An appellate term is an intermediate appellate court that hears appeals from the inferior courts within their designated counties or judicial districts, and are intended to ease the workload on the Appellate Division and provide a less expensive forum closer to the people.', 'As verbs the difference between request and appeal is that request is to express the need or desire for while appeal is (obsolete) to accuse (someone of something).', 'As nouns the difference between fraction and division is that fraction is a part of a whole, especially a comparatively small part while division is (uncountable) the act or process of dividing anything.', 'As nouns the difference between divide and division is that divide is a thing that divides while division is (uncountable) the act or process of dividing anything.', 'definition and meaning , “A feature is a distinctive characteristic of a good or service that sets it apart from similar items. Customers, however, want a benefit..” The same source also defines objective: “A specific result that a person or system aims to achieve within a time frame and with available resources.', ...]	`[1, 0, 0, 0, 0, ...]`

Loss: LambdaLoss with these parameters:

{
    "weighting_scheme": "sentence_transformers.cross_encoder.losses.lambda_loss.NDCGLoss2PPScheme",
    "k": null,
    "sigma": 1.0,
    "eps": 1e-10,
    "reduction_log": "binary",
    "activation_fn": "torch.nn.modules.linear.Identity",
    "mini_batch_size": 8
}

Training Hyperparameters

Non-Default Hyperparameters

per_device_train_batch_size: 16
num_train_epochs: 1
learning_rate: 2e-05
warmup_steps: 0.1
gradient_accumulation_steps: 4
fp16: True
per_device_eval_batch_size: 16
load_best_model_at_end: True
seed: 12
dataloader_num_workers: 2
dataloader_persistent_workers: True

All Hyperparameters

Click to expand

per_device_train_batch_size: 16
num_train_epochs: 1
max_steps: -1
learning_rate: 2e-05
lr_scheduler_type: linear
lr_scheduler_kwargs: None
warmup_steps: 0.1
optim: adamw_torch_fused
optim_args: None
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
optim_target_modules: None
gradient_accumulation_steps: 4
average_tokens_across_devices: True
max_grad_norm: 1.0
label_smoothing_factor: 0.0
bf16: False
fp16: True
bf16_full_eval: False
fp16_full_eval: False
tf32: None
gradient_checkpointing: False
gradient_checkpointing_kwargs: None
torch_compile: False
torch_compile_backend: None
torch_compile_mode: None
use_liger_kernel: False
liger_kernel_config: None
use_cache: False
neftune_noise_alpha: None
torch_empty_cache_steps: None
auto_find_batch_size: False
log_on_each_node: True
logging_nan_inf_filter: True
include_num_input_tokens_seen: no
log_level: passive
log_level_replica: warning
disable_tqdm: False
project: huggingface
trackio_space_id: None
trackio_bucket_id: None
trackio_static_space_id: None
per_device_eval_batch_size: 16
prediction_loss_only: True
eval_on_start: False
eval_do_concat_batches: True
eval_use_gather_object: False
eval_accumulation_steps: None
include_for_metrics: []
batch_eval_metrics: False
save_only_model: False
save_on_each_node: False
enable_jit_checkpoint: False
push_to_hub: False
hub_private_repo: None
hub_model_id: None
hub_strategy: every_save
hub_always_push: False
hub_revision: None
load_best_model_at_end: True
ignore_data_skip: False
restore_callback_states_from_checkpoint: False
full_determinism: False
seed: 12
data_seed: None
use_cpu: False
accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
parallelism_config: None
dataloader_drop_last: False
dataloader_num_workers: 2
dataloader_pin_memory: True
dataloader_persistent_workers: True
dataloader_prefetch_factor: None
remove_unused_columns: True
label_names: None
train_sampling_strategy: random
length_column_name: length
ddp_find_unused_parameters: None
ddp_bucket_cap_mb: None
ddp_broadcast_buffers: False
ddp_static_graph: None
ddp_backend: None
ddp_timeout: 1800
fsdp: None
fsdp_config: None
deepspeed: None
debug: []
skip_memory_metrics: True
do_predict: False
resume_from_checkpoint: None
warmup_ratio: None
local_rank: -1
prompts: None
batch_sampler: batch_sampler
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}

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	-	0.1341 (-0.4572)	0.0124 (-0.5280)	0.2983 (-0.0268)	0.0067 (-0.4940)	0.1058 (-0.3496)
0.0007	1	2.1925	-	-	-	-	-
0.0667	100	2.0361	-	-	-	-	-
0.1334	200	1.0558	-	-	-	-	-
0.1668	250	-	0.7213 (+0.1300)	0.4535 (-0.0869)	0.3667 (+0.0417)	0.3618 (-0.1389)	0.3940 (-0.0614)
0.2001	300	0.6286	-	-	-	-	-
0.2668	400	0.5332	-	-	-	-	-
0.3335	500	0.5208	0.7468 (+0.1556)	0.4826 (-0.0578)	0.3783 (+0.0533)	0.4192 (-0.0815)	0.4267 (-0.0287)
0.4002	600	0.4892	-	-	-	-	-
0.4669	700	0.4780	-	-	-	-	-
0.5003	750	-	0.7538 (+0.1626)	0.5041 (-0.0363)	0.4069 (+0.0819)	0.3785 (-0.1221)	0.4299 (-0.0255)
0.5336	800	0.4448	-	-	-	-	-
0.6003	900	0.4455	-	-	-	-	-
0.6670	1000	0.4427	0.7611 (+0.1698)	0.4584 (-0.0820)	0.3731 (+0.0481)	0.3098 (-0.1908)	0.3804 (-0.0749)
0.7337	1100	0.4307	-	-	-	-	-
0.8004	1200	0.4269	-	-	-	-	-
0.8338	1250	-	0.7640 (+0.1727)	0.4978 (-0.0426)	0.3887 (+0.0636)	0.3634 (-0.1373)	0.4166 (-0.0388)
0.8671	1300	0.4119	-	-	-	-	-
0.9338	1400	0.4104	-	-	-	-	-
1.0	1500	0.4018	0.7664 (+0.1751)	0.5023 (-0.0381)	0.3743 (+0.0492)	0.3538 (-0.1468)	0.4101 (-0.0452)
-1	-1	-	0.7664 (+0.1751)	0.5023 (-0.0381)	0.3743 (+0.0492)	0.3538 (-0.1468)	0.4101 (-0.0452)

The bold row denotes the saved checkpoint.

Training Time

Training: 2.2 hours

Framework Versions

Python: 3.12.13
Sentence Transformers: 5.6.0
Transformers: 5.13.1
PyTorch: 2.11.0+cu128
Accelerate: 1.14.0
Datasets: 4.0.0
Tokenizers: 0.22.2

Additional Resources

Training and Finetuning Reranker Models with Sentence Transformers: the end-to-end guide for training or finetuning Cross Encoder (reranker) models.
Multimodal Embedding & Reranker Models with Sentence Transformers: use text, image, audio, and video reranker models through the same API.
Training and Finetuning Multimodal Embedding & Reranker Models with Sentence Transformers: training multimodal Cross Encoders.

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",
}

LambdaLoss

@inproceedings{wang2018lambdaloss,
  title={The LambdaLoss Framework for Ranking Metric Optimization},
  author={Wang, Xuanhui and Li, Cheng and Golbandi, Nadav and Bendersky, Michael and Najork, Marc},
  booktitle={Proceedings of the 27th ACM international conference on information and knowledge management},
  pages={1313--1322},
  year={2018}
}