teaislife/predrel-nli-deberta-v3

Predicate Alignment DeBERTa-v3

Fine-tuned DeBERTa-v3-base for aligning logical predicates in a neuro-symbolic first-order logic (FOL) pipeline. This model classifies the logical relation between two predicate definitions.

Labels

ID	Label	Meaning
0	ENTAILMENT	Logical implication
1	UNRELATED	No logical relation between the predicates
2	COMPLEMENTARY	Strict negation (A and not-A)

Training data

The model was fine-tuned on a combination of:

• HANS — entailment and unrelated examples for binary predicates, with swapped variables or constants.
• Negation templates — complementary pairs (A / non-A)
• ScoNe-NLI — scoped negation distractors

Usage

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

model_id = "teaislife/predrel-nli-deberta-v3"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSequenceClassification.from_pretrained(model_id)
model.eval()

premise = "The first entity owns the second entity"
hypothesis = "The first entity does not own the second entity"

inputs = tokenizer(premise, hypothesis, return_tensors="pt", truncation=True)
with torch.inference_mode():
    logits = model(**inputs).logits

pred_id = logits.argmax(-1).item()
print(model.config.id2label[pred_id])

Part of

This model is part of a neuro-symbolic pipeline that translates natural-language DAGs into first-order logic and builds bridge axioms between predicates.