Model hallucinating on simple tasks
Just asked the model to read a file and it hallucinated for about an hour until I stopped it.
The files was CURSOR.md and it tried to read CURCurl.md
did it really? is this in OpenClaude or something? I have a feeling that framework's is a bit too much for this model, perhaps a simpler one like Cline/Continue/Kilo/Roo/OpenCode/Codex would be a bit better. In my little experience the claude code source has like 72k or something tokens in it's prompt, how much context were u loading the model with, did you unify kv cache? There have also been major updates in the v2 version of this model, could you please try your test again?
Its claude :)
Yep I saw that haha, just confirming his settings. Also sorry for the late reply to this thread, been swamped with bugs and fixes since the gemma 4 release
did it really? is this in OpenClaude or something? I have a feeling that framework's is a bit too much for this model, perhaps a simpler one like Cline/Continue/Kilo/Roo/OpenCode/Codex would be a bit better. In my little experience the claude code source has like 72k or something tokens in it's prompt, how much context were u loading the model with, did you unify kv cache? There have also been major updates in the v2 version of this model, could you please try your test again?
No, is Claude Code. I just tested it with Ollama to see how it goes before firing up llama.cpp
Please redownload (v2) and test with that model
TeichAI/gemma-4-26B-A4B-it-Claude-Opus-Distill-v2-GGUF works without hallucinating and uses tools successfully as long as I tested it.
TeichAI/gemma-4-26B-A4B-it-Claude-Opus-Distill-v2-GGUFworks without hallucinating and uses tools successfully as long as I tested it.
I'm currently trying to use it as a local agent with Claude Code CLI + Claude Code Router (CCR) + LM Studio (0.4.11), but I keep running into tool calling formatting errors and it fails to create any files.
For example, when it tries to use the Bash or TaskCreate tools, it outputs the wrong data types (like passing an array instead of a string) or hallucinates parameters (like inventing a task_id while missing the required description), causing Claude Code to crash with InputValidationError. It also occasionally leaks raw tokens like } <tool_call|>.