Update README.md

README.md

@@ -37,6 +37,16 @@ We found that [the "value retrieval" variant of NIAH](https://github.com/Arize-a
 
 This model does achieve 100% all-green results on value retrieval but less than satisfactory results on the original version.
 
+### Reproduce
+
+[360k/niah](https://github.com/Qihoo360/360zhinao/blob/main/360k/niah/) generates the raw results.
+
+The score for value retrieval NIAH is calculated on-the-fly when generating the raw results, while the actual score of original and Chinese NIAH is calculated in [360k/plot](https://github.com/Qihoo360/360zhinao/blob/main/360k/plot/).
+
+For the original version, 100% score is given if the regular expression `sandwich.+?dolores.+?sunny` matches the model output, and edit distance otherwise.
+
+For the Chinese version, 100% score is given if `刘秀` is present in the model output, and edit distance otherwise. For the English-biased llama3 models this may not be perfect.
+
 
 ## Usage