dispatchAI/error-patterns

1[2  {3    "error": "Model produces gibberish",4    "cause": "Wrong chat_format",5    "fix": "Set chat_format parameter (llama-3 for SmolLM/Llama, chatml for Qwen, gemma for Gemma)"6  },7  {8    "error": "Model repeats itself",9    "cause": "Small model limitation",10    "fix": "Use a larger model or lower temperature (0.3)"11  },12  {13    "error": "Model ignores instructions",14    "cause": "Raw text completion instead of chat",15    "fix": "Use create_chat_completion() not __call__()"16  },17  {18    "error": "Out of memory on phone",19    "cause": "Model too large for RAM",20    "fix": "Use Q4_K_M quant or smaller model (SmolLM2-135M = 101MB)"21  },22  {23    "error": "Very slow inference",24    "cause": "Large model on weak hardware",25    "fix": "Use 135M or 500M model, or Q4 quantization"26  },27  {28    "error": "Model loads but no output",29    "cause": "Interactive mode in llama-cli",30    "fix": "Add -st flag for single-turn mode"31  },32  {33    "error": "ADB push fails on Windows",34    "cause": "MSYS path conversion",35    "fix": "Use MSYS_NO_PATHCONV=1 before adb command"36  },37  {38    "error": "Arabic text garbled",39    "cause": "Tokenizer not supporting Arabic",40    "fix": "Use Gemma-2B-Arabic with gemma chat format"41  },42  {43    "error": "Function calling fails",44    "cause": "Model not fine-tuned for FC",45    "fix": "Use strong system prompt with function definitions"46  },47  {48    "error": "Model says 'I don't know'",49    "cause": "Small model knowledge limit",50    "fix": "Use larger model (1B+) for factual questions"51  }52]