This dataset is the LLaMini from MBZUAI/LaMini-instruction, removing instances of blatant alignment and removes duplicates.
2290278 instructions remain.
i merged the parquets from original repo with parquet2json then ran clean_format_dedupe.py on the resulting jsonl
credit to ehardford for his contains_unwanted_words function
