metadata

library_name: transformers
tags:
  - natural-language-inference
  - nli
license: mit
datasets:
  - nyu-mll/multi_nli
language:
  - en
base_model: microsoft/deberta-v3-base

deberta v3 base - Natural Language Inference

Model overview

This model is trained for the Natural Language Inference task. It takes two sentences as input (a premise and a hypothesis) and predicts the relationship between them by assigning one of three labels: "entailment," "neutral," or "contradiction." The model is based on the microsoft/deberta-v3-base model, fine-tuned on the nyu-mll/multi_nli dataset, and returns scores corresponding to the labels.

Results

After fine-tuning on the dataset, the model achieved the following results:

Loss: 0.276
Accuracy: 0.899
F1-Score: 0.899

These metrics were evaluated on the validation_mismatched split of the dataset.

Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

model_name = "chincyk/deberta-v3-base-nli"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

premise = "The flight arrived on time at the airport."
hypothesis = "The flight was delayed by several hours."
inputs = tokenizer(premise, hypothesis, return_tensors='pt')

with torch.no_grad():
    outputs = model(**inputs)

logits = outputs.logits
probs = torch.softmax(logits, dim=-1).squeeze()

id2label = model.config.id2label

for i, prob in enumerate(probs):
    print(f"{id2label[i]}: {prob:.4f}")