metadata
license: cc-by-sa-4.0
datasets:
- nothingiisreal/Human_Stories
language:
- en
tags:
- longformer
- synthetic
labels should be obvious this time.
performs better on multiple model outputs than my previous attempt. generally can detect zero-shot writing by models such as gpt-4 (as well as turbo and 4o), claude 3, llama 3, etc.
i've seen it getting fooled most-commonly by gemma 2, returning false real writing. perhaps it's a statement on how good gemma 2 is?