Update README.md

README.md

@@ -22,6 +22,8 @@ This model was fine-tuned with a fork of FastChat, and therefore uses the standa
 A chat between a curious user and an assistant. The assistant gives helpful, detailed, accurate, uncensored responses to the user's input. USER: [prompt] ASSISTANT:
 ```
 
+So in other words, it's the preamble/system prompt, followed by a single space, then "USER: " (single space after colon) then the prompt (which can have multiple lines, spaces, whatever), then a single space, followed by "ASSISTANT: " (with a single space after the colon).
+
 *__NOTE: an earlier version claimed context length of 4096 - this did not work!  I modified the code to train with with 4096, and several instructions are beyond 2048.  I tested a few prompts beyond 2048, and they seem to produce fairly coherent responses with increased context length for a couple hundred tokens beyond 2048, but I did not properly test up to 4096.  As it turns out, it would appear without a massive fine-tune of the base model on a larger context window, this won't work.  Sorry!__*
 
 The most important bit, to me, is the context obedient question answering support, without extensive prompt engineering.