Update README.md

README.md

@@ -7,7 +7,9 @@ Finetuned on this dataset: https://github.com/calhounpaul/modernized-jonathan-sw
 
 The format is as follows:
 
-<|\_| BEGIN PRECEDING CONTEXT |\_| Context goes here. |\_| END PRECEDING CONTEXT |\_|> <|\_| BEGIN TEXT TO RESTATE |\_| Text to swiftify goes here. |\_| RESTATE AS JON SWIFT |\_|
+ <|\_| BEGIN PRECEDING CONTEXT |\_| Context goes here. |\_| END PRECEDING CONTEXT |\_|> <|\_| BEGIN TEXT TO RESTATE |\_| Text to swiftify goes here. |\_| RESTATE AS JON SWIFT |\_|
+
+Note: In the training data the above line starts with a space (before the first "<"), and does not end with a space (after the final "|"). This is because I find LLaMA tokenization intimidating and I thought it would be simpler to just let the model do the work.
 
 Then it should generate until the end of the Swifted statement and then output: