error stringclasses 10
values | cause stringclasses 10
values | fix stringclasses 10
values |
|---|---|---|
Model produces gibberish | Wrong chat_format | Set chat_format parameter (llama-3 for SmolLM/Llama, chatml for Qwen, gemma for Gemma) |
Model repeats itself | Small model limitation | Use a larger model or lower temperature (0.3) |
Model ignores instructions | Raw text completion instead of chat | Use create_chat_completion() not __call__() |
Out of memory on phone | Model too large for RAM | Use Q4_K_M quant or smaller model (SmolLM2-135M = 101MB) |
Very slow inference | Large model on weak hardware | Use 135M or 500M model, or Q4 quantization |
Model loads but no output | Interactive mode in llama-cli | Add -st flag for single-turn mode |
ADB push fails on Windows | MSYS path conversion | Use MSYS_NO_PATHCONV=1 before adb command |
Arabic text garbled | Tokenizer not supporting Arabic | Use Gemma-2B-Arabic with gemma chat format |
Function calling fails | Model not fine-tuned for FC | Use strong system prompt with function definitions |
Model says 'I don't know' | Small model knowledge limit | Use larger model (1B+) for factual questions |
- Downloads last month
- 7