The english_only subset is taken from the downsampled configuration of agentlans/en-chat-refusal.
The multilingual subset is created by mixing agentlans/multilingual-chat-refusal with the english_only subset in a 1:1 ratio.
All conversations have been formatted using role tokens and ellipses <|...|> for long texts.
The resulting files are shuffled and split 80% training, 20% testing.
