Gltr.Io

What is Gltr.Io

GLTR is a tool developed by MIT-IBM Watson AI Lab and HarvardNLP to detect automatically generated text. It allows for a forensic analysis of the visual footprint of language models, determining the likelihood of a text being generated by an automatic system. By analyzing the ranking and colors of words in the text, GLTR provides a visual indication of how likely each word was generated by a language model. Despite some limitations, GLTR can serve as a starting point for the development of similar ideas at a larger scale.

Key Features

• Forensic Analysis: Detects automatically generated text.
• Language Model Insights: Understands how large models produce text.
• Visual Indication: Color-codes words based on prediction likelihood.
• Histograms: Shows word category distributions and prediction entropies.
• GPT-2 117M Language Model: Access to one of the largest publicly available models.
• Positional Analysis: Overlay colored masks to indicate word likelihood.
• Histograms and Statistics: Display word category distribution, prediction probabilities, and entropies.
• Real vs Generated Text Analysis: Detect artificial text by comparing predictability and uncertainty.