metadata
license: cc-by-4.0
tags:
- sharegpt
- reasoning
- chat
- safety
size_categories:
- 1K<n<10K
DONOT BS
DONOT BS is a dataset designed to train LLM's to avoid topics which could cause hallucinations, and simply reject. The dataset is multi turn, every turn has a fake user agent which is trying to cause the LLM to not reject.
The dataset includes ~4000+ rows, all in a sharegpt format with reasoning between think tags.
first turn prompts are cc-by-4.0, llm output under apache 2.0, second turn prompts under apache 2.0.
models used: (no specific order)
- laguna xs 2.1
- muse glimmer
- gemma 4 26b a4b