Update README.md

README.md

@@ -11,6 +11,12 @@ inference: false
 <h1 style="text-align: center">Metharme 7B</h1>
 <h2 style="text-align: center">An instruction-tuned LLaMA biased towards fiction writing and conversation.</h2>
 
+> Currently KoboldCPP is unable to stop inference when an EOS token is emitted, which causes the model to devolve into gibberish,
+> 
+> Metharme 7B is now fixed on the dev branch of KoboldCPP, which has fixed the EOS issue. Make sure you're compiling the latest version, it was fixed only a after this model was released;
+> 
+> When running KoboldCPP, you will need to add the --unbantokens flag for this model to behave properly. 
+
 ## Model Details
 
 This models has the XOR files pre-applied out of the box.