Training dataset from our paper Large-Scale RLVR Improves Instruction Hierarchy on Frontier LLMs.
To avoid legal and reputational risk, we replaced all company names in the original dataset for either COMPETITOR_i or BRAND_i, with i ∈ ℕ. We recommend that you replace these placeholders with real company names before training on the dataset.
attacker_meta_problem
str
General… See the full description on the dataset page:
https://huggingface.co/datasets/openai/ih-challenge.