Factual Consistency Evaluator/Metric in ACL 2023 paper

WeCheck: Strong Factual Consistency Checker via Weakly Supervised Learning

Model description

WeCheck is a factual consistency metric trained from weakly annotated samples.

This WeCheck checkpoint can be used to check the following three generation tasks:

Text Summarization/Knowlege grounded dialogue Generation/Paraphrase

This WeCheck checkpoint is trained based on the following three weak labler:

QAFactEval / Summarc / NLI warmup

How to use the model

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
model_name = "nightdessert/WeCheck"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
premise = "I first thought that I liked the movie, but upon second thought it was actually disappointing." # Input for Summarization/ Dialogue / Paraphrase
hypothesis = "The movie was not good."   # Output for Summarization/ Dialogue / Paraphrase
input = tokenizer(premise, hypothesis, truncation=True, return_tensors="pt", truncation_strategy="only_first", max_length=512)
output = model(input["input_ids"].to(device))['logits'][:,0]  # device = "cuda:0" or "cpu"
prediction = torch.sigmoid(output).tolist()
print(prediction)

license: openrail pipeline_tag: text-classification tags: - Factual Consistency - Natrual Language Inference

language:

en tags:
Factual Consistency Evaluation