Update README.md

README.md

@@ -11,4 +11,11 @@ PontAvignon is a specialized reasoning model using a special tokens to encode in
 
 Any excerpts of a show program has to be submitted using the tags '<|text_startl>' and '<|text_end|>'.
 
+After '<|text_end|>' will model will generate an answer in three step:
+* Initial fuzzy "thinking" (between <|thinking_start|> and <|thinking_end|>): 
+
+Some warnings apply:
+* Given the intensive reinforcement learning specialization, the model will mostly work on theater show in French, preferably one structured similarly to original sources from the festival d'Avignon. More diverse sources would be needed to make the model more source agnostic.
+* To improve results accuracy, we trained the model on filtered versions of the original sources with all the information about a unique show. Multi-show submissions will either result on the model focusing on only one show or, even more problematic, mixing information.
+* Reasoning traces contribute to increase the accuracy of the model and provide some level of explainability. They can still be counterintuitive and potentially contain wrong assumptions that will not necessarily be kept in the final output.