First model in the potential series.
<s> [INST] QUERY_1 [/INST] ANSWER_1</s> [INST] QUERY_2 [/INST] ANSWER_2</s>...
Beware that some backends (like llama.cpp) add bos already (by default), so you don't need to prepend it yourself.
DO NOT CHANGE ROPE SETTINGS. This model uses high freq base with 32k seen tokens, it should be fine for most tasks.
Only tested with temp 1 and top_p 0.95 with everything else disabled.