Data decontamination - n-gram matching works better than embedding methods in terms of precision - 8-gram matching only on user-turns (since completions are LLM generated) - remove training sets that have more than 2% overlap with any evaluation set
0 likes 1 replies
?