Tool call format degrading at higher context

#10
by ilintar - opened

So I've actually managed to run the model quite well with OpenCode and it does perform well at context lengths up to ~60k. However, at around ~60k, it suddenly starts performing tool calls in a JSON format, often in the reasoning section as well. Any idea what the issue is and can it maybe be remedied in some post-training?

Hi @ilintar , could you share the raw output so we can take a look? Thanks

Here's the entire conversation dumped into an OAI-compatible message history:

https://gist.github.com/pwilkin/5e5037ecface2789f6cdd612bb29b6ed

ilintar changed discussion status to closed
ilintar changed discussion status to open

Sign up or log in to comment