StepJumpy4782

wow super cool great work. could you explain more on what long document dataset and prompts were? since that is a notable usecase for me and your data showing it performing the worst is interesting to me. how are you running these? among latest qwen you said were working on, could we get some analysis like this for the really big ones (GLM 5.1 just came out :)). anything the community could help with in that regard?

model protocol

meh I don't mind posts like these because im not staying in the loop of llama.cpp changes even though I pull latest daily. what are people doing to stay on top of things? the releases notes are fairly sparse. but yeah you need to relax. just use an older version and wait for the fix or patch it as you did.

lmaoo One thing that comes to mind is kinds of backdoors. Could train it that specific prompts anticipated to be used by the enemy are intentionally really bad / include obscure vulns / bad advise etc. Now if thats actually happening, well remains to be seen. I certainly have my doubts. Its open after all and a finding like that would instantly destroy so much trust they would have built up.

A bit of the loop with the latest happenings, will give 4.7 a go. What specifically makes GLM 4.7 stand out compared to everyone else? What more can we expect with future releases (closed and open)? And more specifically, what future areas of research are you guys most interesting in learning about?