This is a Cross Encoder model finetuned from cross-encoder/nli-deberta-v3-xsmall using the sentence-transformers library. It computes scores for pairs of texts, which can be used for text pair classification.

Model Details

Model Description

Model Type: Cross Encoder
Base model: cross-encoder/nli-deberta-v3-xsmall
Maximum Sequence Length: 512 tokens
Number of Output Labels: 3 labels
Supported Modality: Text

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': 'DebertaV2ForSequenceClassification'})
)

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.

python

from sentence_transformers import CrossEncoder

# Download from the 🤗 Hub
model = CrossEncoder("kelompoknlp2026dsindo/deberta-v3-xsmall-nli-finetuned")
# Get scores for pairs of inputs
pairs = [
    ['[South Australia] has the most expensive electricity in the world.', '[citation needed] South Australia has the highest retail price for electricity in the country.'],
    ['[South Australia] has the most expensive electricity in the world.', '"South Australia has the highest power prices in the world".'],
    ['when 3 per cent of total annual global emissions of carbon dioxide are from humans and Australia prod\xaduces 1.3 per cent of this 3 per cent, then no amount of emissions reductio\xadn here will have any effect on global climate.', 'The 2011 UNEP Green Economy report states that "[a]agricultural operations, excluding land use changes, produce approximately 13 per cent of anthropogenic global GHG emissions.'],
    ['when 3 per cent of total annual global emissions of carbon dioxide are from humans and Australia prod\xaduces 1.3 per cent of this 3 per cent, then no amount of emissions reductio\xadn here will have any effect on global climate.', 'With a market share of 30% and (potentially) clean electricity, heat pumps could reduce global CO 2 emissions by 8% annually.'],
    ['when 3 per cent of total annual global emissions of carbon dioxide are from humans and Australia prod\xaduces 1.3 per cent of this 3 per cent, then no amount of emissions reductio\xadn here will have any effect on global climate.', 'In the modern era, emissions to the atmosphere from volcanoes are approximately 0.645 billion tonnes of CO 2 per year, whereas humans contribute 29 billion tonnes of CO 2 each year.'],
]
scores = model.predict(pairs)
print(scores)
# [[-5.1375  3.7547 -0.2361]
#  [-4.5641  3.9137 -1.1062]
#  [ 1.3195 -2.9018  1.4899]
#  [ 0.1377 -2.4617  2.3478]
#  [ 1.4282 -2.7619  1.3155]]

Evaluation

Metrics

Cross Encoder Classification

Datasets: DeBERTa-xsmall-finetune-dev, Claim-Evidence-train and Claim-Evidence-3Class-dev
Evaluated with CrossEncoderClassificationEvaluator

Metric	DeBERTa-xsmall-finetune-dev	Claim-Evidence-train	Claim-Evidence-3Class-dev
f1_macro	0.7113	0.9316	0.713
f1_micro	0.7344	0.9287	0.7367
f1_weighted	0.7344	0.9287	0.7366

Training Details

Training Dataset

Unnamed Dataset

Size: 3,730 training samples
Columns: claim, evidence, and label
Approximate statistics based on the first 1000 samples:
claim evidence label
type string string int
details
min: 8 tokens
mean: 25.8 tokens
max: 64 tokens
min: 6 tokens
mean: 37.91 tokens
max: 459 tokens
0: ~11.10%
1: ~35.90%
2: ~53.00%

Samples:

claim	evidence	label
`El Niño drove record highs in global temperatures suggesting rise may not be down to man-made emissions.`	`While ‘climate change’ can be due to natural forces or human activity, there is now substantial evidence to indicate that human activity – and specifically increased greenhouse gas (GHGs) emissions – is a key factor in the pace and extent of global temperature increases.`	`0`
`El Niño drove record highs in global temperatures suggesting rise may not be down to man-made emissions.`	`This acceleration is due mostly to human-caused global warming, which is driving thermal expansion of seawater and the melting of land-based ice sheets and glaciers.`	`0`
`In 1946, PDO switched to a cool phase.`	`There is evidence of reversals in the prevailing polarity (meaning changes in cool surface waters versus warm surface waters within the region) of the oscillation occurring around 1925, 1947, and 1977; the last two reversals corresponded with dramatic shifts in salmon production regimes in the North Pacific Ocean.`	`1`

Loss: CrossEntropyLoss

Evaluation Dataset

Unnamed Dataset

Size: 433 evaluation samples
Columns: claim, evidence, and label
Approximate statistics based on the first 433 samples:
claim evidence label
type string string int
details
min: 7 tokens
mean: 27.42 tokens
max: 76 tokens
min: 8 tokens
mean: 35.4 tokens
max: 217 tokens
0: ~13.16%
1: ~39.49%
2: ~47.34%

Samples:

claim	evidence	label
`[South Australia] has the most expensive electricity in the world.`	`[citation needed] South Australia has the highest retail price for electricity in the country.`	`1`
`[South Australia] has the most expensive electricity in the world.`	`"South Australia has the highest power prices in the world".`	`1`
`when 3 per cent of total annual global emissions of carbon dioxide are from humans and Australia produces 1.3 per cent of this 3 per cent, then no amount of emissions reduction here will have any effect on global climate.`	`The 2011 UNEP Green Economy report states that "[a]agricultural operations, excluding land use changes, produce approximately 13 per cent of anthropogenic global GHG emissions.`	`2`

Loss: CrossEntropyLoss

Training Hyperparameters

Non-Default Hyperparameters

per_device_train_batch_size: 16
gradient_accumulation_steps: 2
learning_rate: 3e-05
weight_decay: 0.05
num_train_epochs: 5
lr_scheduler_type: cosine
warmup_steps: 0.1
load_best_model_at_end: True

All Hyperparameters

Click to expand

do_predict: False
prediction_loss_only: True
per_device_train_batch_size: 16
per_device_eval_batch_size: 8
gradient_accumulation_steps: 2
eval_accumulation_steps: None
torch_empty_cache_steps: None
learning_rate: 3e-05
weight_decay: 0.05
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 5
max_steps: -1
lr_scheduler_type: cosine
lr_scheduler_kwargs: None
warmup_ratio: None
warmup_steps: 0.1
log_level: passive
log_level_replica: warning
log_on_each_node: True
logging_nan_inf_filter: True
enable_jit_checkpoint: False
save_on_each_node: False
save_only_model: False
restore_callback_states_from_checkpoint: False
use_cpu: False
seed: 42
data_seed: None
bf16: False
fp16: False
bf16_full_eval: False
fp16_full_eval: False
tf32: None
local_rank: -1
ddp_backend: None
debug: []
dataloader_drop_last: False
dataloader_num_workers: 0
dataloader_prefetch_factor: None
disable_tqdm: False
remove_unused_columns: True
label_names: None
load_best_model_at_end: True
ignore_data_skip: False
fsdp: []
fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': 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
deepspeed: None
label_smoothing_factor: 0.0
optim: adamw_torch_fused
optim_args: None
group_by_length: False
length_column_name: length
project: huggingface
trackio_space_id: trackio
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
push_to_hub: False
resume_from_checkpoint: None
hub_model_id: None
hub_strategy: every_save
hub_private_repo: None
hub_always_push: False
hub_revision: None
gradient_checkpointing: False
gradient_checkpointing_kwargs: None
include_for_metrics: []
eval_do_concat_batches: True
auto_find_batch_size: False
full_determinism: False
ddp_timeout: 1800
torch_compile: False
torch_compile_backend: None
torch_compile_mode: None
include_num_input_tokens_seen: no
neftune_noise_alpha: None
optim_target_modules: None
batch_eval_metrics: False
eval_on_start: False
use_liger_kernel: False
liger_kernel_config: None
eval_use_gather_object: False
average_tokens_across_devices: True
use_cache: False
prompts: None
batch_sampler: batch_sampler
multi_dataset_batch_sampler: proportional
router_mapping: {}
learning_rate_mapping: {}

Training Logs

Epoch	Step	Training Loss	Validation Loss	DeBERTa-xsmall-finetune-dev_f1_macro	Claim-Evidence-train_f1_macro	Claim-Evidence-3Class-dev_f1_macro
0.4274	50	1.6490	0.9456	0.5752	-	-
0.8547	100	0.8194	0.7744	0.6204	-	-
1.2821	150	0.6911	0.8904	0.6485	-	-
1.7094	200	0.6048	0.8325	0.6564	-	-
2.1368	250	0.5231	0.8821	0.6783	-	-
2.5641	300	0.3898	0.9131	0.6917	-	-
2.9915	350	0.3996	0.9402	0.6767	-	-
3.4188	400	0.2743	0.9794	0.7206	-	-
3.8462	450	0.3070	1.0230	0.7115	-	-
4.2735	500	0.2496	1.0124	0.7101	-	-
4.7009	550	0.2757	1.0294	0.7113	-	-
-1	-1	-	-	-	0.9316	0.7130

Training Time

Training: 14.3 minutes

Framework Versions

Python: 3.12.13
Sentence Transformers: 5.4.1
Transformers: 5.0.0
PyTorch: 2.10.0+cu128
Accelerate: 1.13.0
Datasets: 4.0.0
Tokenizers: 0.22.2

Citation

BibTeX

Sentence Transformers

bibtex

@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",
}

	claim	evidence	label
type	string	string	int
details	min: 8 tokens mean: 25.8 tokens max: 64 tokens	min: 6 tokens mean: 37.91 tokens max: 459 tokens	0: ~11.10% 1: ~35.90% 2: ~53.00%