tasksource/ModernBERT-base-nli

Model Card for Model ID

ModernBERT multi-task fine-tuned on tasksource NLI tasks, including MNLI, ANLI, SICK, WANLI, doc-nli, LingNLI, FOLIO, FOL-NLI, LogicNLI, Label-NLI and all datasets in the below table). This is the equivalent of an "instruct" version.

Test accuracy at 100k training steps. 215k steps version coming december 24th.

test_name	test_accuracy
glue/mnli	0.91
glue/qnli	0.93
glue/rte	0.86
super_glue/cb	0.89
anli/a1	0.62
anli/a2	0.47
anli/a3	0.42
sick/label	0.92
sick/entailment_AB	0.84
snli	0.91
scitail/snli_format	0.95
hans	1
WANLI	0.71
recast/recast_sentiment	0.98
recast/recast_verbcorner	0.94
recast/recast_ner	0.87
recast/recast_factuality	0.93
recast/recast_puns	0.93
recast/recast_kg_relations	0.94
recast/recast_verbnet	0.88
recast/recast_megaveridicality	0.87
probability_words_nli/usnli	0.77
probability_words_nli/reasoning_1hop	0.99
probability_words_nli/reasoning_2hop	0.9
nan-nli	0.85
nli_fever	0.72
breaking_nli	1
conj_nli	0.71
fracas	0.86
dialogue_nli	0.88
mpe	0.73
dnc	0.9
recast_white/fnplus	0.81
recast_white/sprl	0.92
recast_white/dpr	0.61
robust_nli/IS_CS	0.76
robust_nli/LI_LI	0.98
robust_nli/ST_WO	0.85
robust_nli/PI_SP	0.74
robust_nli/PI_CD	0.8
robust_nli/ST_SE	0.78
robust_nli/ST_NE	0.86
robust_nli/ST_LM	0.81
robust_nli_is_sd	1
robust_nli_li_ts	0.91
add_one_rte	0.91
cycic_classification	0.83
lingnli	0.82
monotonicity-entailment	0.95
scinli	0.79
naturallogic	0.91
syntactic-augmentation-nli	0.95
autotnli	0.92
defeasible-nli/atomic	0.76
defeasible-nli/snli	0.79
help-nli	0.91
nli-veridicality-transitivity	0.99
lonli	0.99
dadc-limit-nli	0.67
folio	0.59
tomi-nli	0.53
temporal-nli	0.92
counterfactually-augmented-snli	0.74
cnli	0.81
logiqa-2.0-nli	0.57
mindgames	0.94
ConTRoL-nli	0.65
logical-fallacy	0.31
conceptrules_v2	0.99
zero-shot-label-nli	0.74
scone	0.97
monli	0.98
SpaceNLI	1
propsegment/nli	0.91
SDOH-NLI	1
scifact_entailment	0.78
AdjectiveScaleProbe-nli	0.99
resnli	0.99
semantic_fragments_nli	0.99
dataset_train_nli	0.88
ruletaker	0.91
PARARULE-Plus	1
logical-entailment	0.73
nope	0.54
LogicNLI	0.65
contract-nli/contractnli_a/seg	0.87
contract-nli/contractnli_b/full	0.78
nli4ct_semeval2024	0.6
biosift-nli	0.88
SIGA-nli	0.54
FOL-nli	0.71
doc-nli	0.82
mctest-nli	0.89
idioms-nli	0.86
lifecycle-entailment	0.71
MSciNLI	0.82
hover-3way/nli	0.9
seahorse_summarization_evaluation	0.82
babi_nli	0.94
gen_debiased_nli	0.9

Usage

[ZS] Zero-shot classification pipeline

from transformers import pipeline
classifier = pipeline("zero-shot-classification",model="tasksource/ModernBERT-base-nli")

text = "one day I will see the world"
candidate_labels = ['travel', 'cooking', 'dancing']
classifier(text, candidate_labels)

NLI training data of this model includes label-nli, a NLI dataset specially constructed to improve this kind of zero-shot classification.

[NLI] Natural language inference pipeline

from transformers import pipeline
pipe = pipeline("text-classification",model="tasksource/ModernBERT-base-nli")
pipe([dict(text='there is a cat',
  text_pair='there is a black cat')]) #list of (premise,hypothesis)

Backbone for further fune-tuning

This checkpoint has stronger reasoning and fine-grained abilities than the base version and can be used for further fine-tuning.

Citation

@inproceedings{sileo-2024-tasksource,
    title = "tasksource: A Large Collection of {NLP} tasks with a Structured Dataset Preprocessing Framework",
    author = "Sileo, Damien",
    booktitle = "Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)",
    month = may,
    year = "2024",
    address = "Torino, Italia",
    publisher = "ELRA and ICCL",
    url = "https://aclanthology.org/2024.lrec-main.1361",
    pages = "15655--15684",
}