Gibberish, odd symbols or ignoring instructions
Symptom
Output contains special tokens, repeats role names, ignores the system prompt or switches language randomly.
Likely causes
- Wrong or missing chat template
- Using a base model instead of the instruct variant
- Over-aggressive quantization (Q2/Q3)
- Rope/context settings changed incorrectly
How to fix it
- Confirm you downloaded the instruct/chat variant.
- Use the model's built-in template:
--jinjain llama-server, or the right template in Ollama. - Try a higher quant (Q5_K_M or Q6_K) to rule out quantization damage.
- Reset sampling to the model card's recommended values.
- Don't override context scaling (rope) unless required.
bash
llama-server -m model.gguf --jinja -ngl 99 -c 8192Still stuck? Check the lessons or measure first with the benchmark lab.