license: mit language: en pipeline_tag: text-classification tags:
- fact-checking
- fake-news
- deberta
- credo-ai
fact-check1-v3-final (Credo AI - Brain 2)
This is the definitive "Brain 2" model from the Credo AI project, a powerful and robust binary (FAKE/REAL) news classifier. This model represents the culmination of a multi-stage refinement process designed to create a highly accurate and unbiased fake news specialist.
The Model's Journey: From Specialist to Sage
This model is a fine-tuned version of microsoft/deberta-v3-large.
v1: The model was initially trained on a massive, combined corpus of over 50,000 news articles, achieving 99.9%+ accuracy on its core task. However, this intense specialization made it a "conspiracy theorist," classifying simple, factual statements as FAKE because they didn't match the stylistic patterns of the news it was trained on.
v2: The model underwent a "scientific calibration" by being fine-tuned on a dataset of pristine scientific facts (
allenai/scifact). This corrected its bias against clean, encyclopedic text but was still too narrow.v3 (This version): The model has completed its final "masterclass." It was fine-tuned on a perfectly balanced dataset containing:
- General Knowledge: Thousands of clean, factual statements from Wikipedia (
wiki_qadataset) to broaden its understanding of truth. - Memory Reinforcement: A sample of classic fake news articles (
mrisdal/fake-news) to remind it of its core mission and prevent catastrophic forgetting.
- General Knowledge: Thousands of clean, factual statements from Wikipedia (
This final calibration, performed with an ultra-low learning rate, has produced a model that is both a powerful fake news detector and a robust, general-purpose fact-checker.
Intended Use
This model is intended to be used as a fast and powerful first-line filter for English-language news headlines and short texts.
from transformers import pipeline
classifier = pipeline("text-classification", model="Arko007/fact-check1-v3-final")
# Example 1: Simple Fact (The test it used to fail)
text1 = "The sun rises in the east."
print(classifier(text1))
# Expected output: [{'label': 'REAL', 'score': ...}]
# Example 2: Fake News
text2 = "BREAKING: Scientists confirm lizards are secretly running the government."
print(classifier(text2))
# Expected output: [{'label': 'FAKE', 'score': ...}]
Limitations
While this v3 model is far more robust, its core training was on political and general news. Its performance may be lower on highly specialized domains like financial analysis or deep scientific literature.