darkc0de/XORTRON.CriminalComputing.Config.LARGE.XPRT2
darkc0de/XORTRON.CriminalComputing.Config.LARGE.XPRT2
mistralai/Mistral-Medium-3.5-128B
Hmm for some reason XORTRON.CriminalComputing.Config.LARGE.XPRT2 gives me broken outputs like
88q the conclusion is a certainizelget the result of \zb \ung(\z }#{ I have my problem, bending theilbenf on the conclusion,$i8um0erned $In this $$ ( beZbj&[ in^#z \newiK>UJ to a +Z; we havez in the years of theaKj thenZB, "ung+6q UVbushies<bbbj^J[5D I can4 IJ88n this possible>nze#}$bjn>7Pzw) {
web9atz);12+ to begin theFZ\nch? given Z#{\ }b- is in theb $Dn2rna0}$;71 die j^{\ Pby8z%l with each die###0aJ= $8B [Sq Lu with{i);
The Transformers config on the original Mistral-Medium-3.5-128B had an incorrect entry that caused long-context performance degradation. Mistral released a fix shortly after my heretic run so that maybe the issue. I've been meaning to rerun with heretic 1.3.0 on the patched commit but haven't gotten to it yet.
It's good to hear from you, was starting to think you might have disappeared into the abyss