# Answer-only runtime package

This contains the already-converted Q4_K_M GGUF and the answer-only runtime preset. The weights are unchanged. It is not a separately fine-tuned answer-only model.

For Ollama, run from this folder:

```sh
ollama create climate-answer-only -f Modelfile
ollama run climate-answer-only
```

For Flutter or another GGUF host, load the GGUF and explicitly apply the system prompt, stop sequences, output limit and temperature in `runtime-settings.json`. This JSON is an application-owned settings file, not a standard GGUF feature; runtimes will not automatically read it. Use the same system prompt when counting context tokens and generating.

Stop before displaying any of the configured stop strings. If the inference engine does not support string stops, buffer partial matches across streamed chunks before displaying text, then cancel when a full stop marker is matched. Never simply remove the labels and display the following invented attribution. Exact stop strings can miss differently formatted citations, so this remains an output-control measure rather than a factual fix.

GGUF preserves weights and tokenizer metadata; the Ollama SYSTEM and PARAMETER directives are not automatically embedded as portable behavior. A model trained without citation targets would require a new training run. This package does not remove the citation training from the weights or guarantee source-free answers.