[
  {
    "error": "Model produces gibberish",
    "cause": "Wrong chat_format",
    "fix": "Set chat_format parameter (llama-3 for SmolLM/Llama, chatml for Qwen, gemma for Gemma)"
  },
  {
    "error": "Model repeats itself",
    "cause": "Small model limitation",
    "fix": "Use a larger model or lower temperature (0.3)"
  },
  {
    "error": "Model ignores instructions",
    "cause": "Raw text completion instead of chat",
    "fix": "Use create_chat_completion() not __call__()"
  },
  {
    "error": "Out of memory on phone",
    "cause": "Model too large for RAM",
    "fix": "Use Q4_K_M quant or smaller model (SmolLM2-135M = 101MB)"
  },
  {
    "error": "Very slow inference",
    "cause": "Large model on weak hardware",
    "fix": "Use 135M or 500M model, or Q4 quantization"
  },
  {
    "error": "Model loads but no output",
    "cause": "Interactive mode in llama-cli",
    "fix": "Add -st flag for single-turn mode"
  },
  {
    "error": "ADB push fails on Windows",
    "cause": "MSYS path conversion",
    "fix": "Use MSYS_NO_PATHCONV=1 before adb command"
  },
  {
    "error": "Arabic text garbled",
    "cause": "Tokenizer not supporting Arabic",
    "fix": "Use Gemma-2B-Arabic with gemma chat format"
  },
  {
    "error": "Function calling fails",
    "cause": "Model not fine-tuned for FC",
    "fix": "Use strong system prompt with function definitions"
  },
  {
    "error": "Model says 'I don't know'",
    "cause": "Small model knowledge limit",
    "fix": "Use larger model (1B+) for factual questions"
  }
]