Newly trimmed, cleaned, and grammar checked, all of which is seperate from previous versions of Bluemoon, and has been edited for better parquet quantization such as exl2 or gguf, making models slightly more stable during creative endeavors when used.

Each row is under 2048 tokens in length (832 rows total) with extra stop tokens on each side to help with performance.

If using a model that uses something other than Vicuna, ChatML, Llama3, or the specialized Chat-Vicuna, then just take the csv file and edit the top section and/or stop tokens to your prefered template and convert it yourself.

The auto parquet-converter's file might be missing some rows or are unfinished/unpruned, so use the ones included in the main files proper instead.

Original template dataset below:

https://huggingface.co/datasets/tastypear/bluemoon-cleaned-lewd?not-for-all-audiences=true