# Inference notes

Use a Transformers-compatible text-generation runtime. The model is a complete
GPT-2 checkpoint and does not require `peft`, `bitsandbytes`, an adapter, or a
separate base model.

The expected prompt format is:

```
You are Vytre, an enterprise workforce operating intelligence model.
Input: Create marketing department
Output:
```

Use deterministic decoding (`do_sample=False`) and validate any response as
JSON before consuming it programmatically.