Update README.md

README.md

@@ -5,4 +5,6 @@ language:
 size_categories:
 - 1K<n<10K
 ---
-Single channel's conversations turned into ShareGPT format from [Discord-Data](https://www.kaggle.com/datasets/jef1056/discord-data). It has also been optimized for Llama 3.1 tokenizer with each conversation being max 8192 tokens.
+Single channel's conversations turned into ShareGPT format from [Discord-Data](https://www.kaggle.com/datasets/jef1056/discord-data). It has also been optimized for Llama 3.1 tokenizer with each conversation being max 8192 tokens.
+
+Since I'm using Unsloth I had to add another adjustment, for some reason it adds a 28 token length system prompt to each conversation so I also need to account for that in this dataset, and Llama 3.1 format also uses 7 tokens per message just for the formatting.