Update README.md

README.md

@@ -19,7 +19,7 @@ The models are offered as GGUFs quantized to Q4.
 * The model can be creative in working around the lack of a letter 'e', for example instead of sheep it can say "adult lamb" or "dominant wool-producing animal".
 * Sometimes it drops the 'e' and mis-spells a word
 * I did not filter out some cyrillic letters which look like and 'e' and sometimes it uses those (or even Chinese words)
-* The smaller model and output refusals if it doesn't find a suitable token
+* The smaller model can output refusals if it doesn't find a suitable token
 * Instead of greedy sampling or probability sampling, beam search with back-tracking is useful to avoid leading the generation into an 'e' dominated dead-end
 
 ### Example