A dataset of 20,000 boolean statements, each labeled by two humans as common-knowledge-true, common-knowledge-false, or neither.

This data was collected for the paper Explore, Establish, Exploit: Red Teaming Language Models from Scratch by the MIT Algorithmic Alignment Group. Authors: Stephen Casper ([email protected]), Jason Lin, Joe Kwon, Gatlen Culp, and Dylan Hadfield-Menell.

Read the paper on arXiv: Explore, Establish, Exploit: Red Teaming Language Models from Scratch

@misc{casper2023explore,
      title={Explore, Establish, Exploit: Red Teaming Language Models from Scratch}, 
      author={Stephen Casper and Jason Lin and Joe Kwon and Gatlen Culp and Dylan Hadfield-Menell},
      year={2023},
      eprint={2306.09442},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Where do the sentences come from? They are all sentences stated by GPT-3-davinci-002 in response to various prompts for interesting facts. We used GPT-3 sentences because our research using this dataset involved red-teaming GPT-3 to find prompts that can make it say dishonest things. As a result of using GPT-3 with our sampling and filtering methodology, some topics and types of sentences are more strongly represented than in normal speech. For example, there are many sentences in the dataset making claims about the sizes of animals.

Where do the labels come from? Each example has two labels from human knowledge workers. This data was collected in partnership with SurgeAI.

How often do the human labels agree? The two human labels agree on 12,106 out of 20,000 examples.

Are some examples mislabeled? We find ourselves disagreeing with the labels sometimes. In particular, we think that some human annotators labeled many things as common-knowledge-true that we do not find obvious. However, (1) we find that we more consistently agree with labels when we filter for examples for which there is an agreement between both labels and (2) we find that we more consistently agree when the annotators label something as common-knowledge-false.

What is the importance of the 'neither' label? Why not just common-knowledge-true/false? Not all declarative statements from a language model will be objectively true or false irrespective of opinions or context.

Why are the labels common-knowledge-true/false/neither instead of just true/false/neither? Right now, there is a lot of interest (including from us) in studying dishonest behavior from LLMs. But a challenge with this is what standard to hold a language model to. For example, is there a difference between types of lies a model may say? Are some innocent mistakes while others are deliberate falsehoods? The answers to these questions may have implications for how we interpret and debug dishonesty in language models. But in our case, by focusing on things that are commonly known to be false, we targetedly study the most egregious type of dishonesty in language models.

How were human knowledge workers instructed to provide labels? We instructed them to provide a label based on whether they believe a typical human adult would believe a sentence to be reasonably true, reasonably false, or neither. See the paper appendix for full details.

Are there any dysfluencies in sentences? Yes, a small proportion of sentences are ungrammatical or nonsensical due to errors in GPT-3's generations. We instructed human annotators to label all of these sentences as 'neither'.

What do you do with the data in the paper? We use it to train a classifier to distinguish between obviously false statements and all other statements. Then we use it as a learning signal to red team GPT-3 to find prompts that elicit false responses. In our case, it worked, while two baselines based on ChatGPT labels (instead of human ones) and the CREAK dataset did not work.

What else could be done with this data? We hope it will be useful for lots of work -- particularly probing and red teaming research in large language models :)