Qwen3-8B NPC dialogue LoRA

LoRA adapter that makes Qwen3-8B play the NPCs of a Godot RPG: it answers in character (Brazilian Portuguese) and returns the JSON the game executes, {"reply", "action", "mood_shift"}.

On a held-out set of 392 turns (unseen worlds and characters), the action matches the reference 92.9% of the time against 10.8% for the base model, with 100% valid JSON. Inside the game, the rule validator approved 97.5% of the replies.

Training code, metrics and the merge-vs-runtime-LoRA comparison: https://github.com/NicoDeVargas/qwen3-npc-dialogue-lora

Files

• adapter_model.safetensors, adapter_config.json: PEFT adapter (r 16, alpha 32, all projections).
• qwen3-8b-npc-dialogue-lora-f16.gguf: the same adapter for llama.cpp.

Use

llama-server -m Qwen3-8B-Q4_K_M.gguf --lora qwen3-8b-npc-dialogue-lora-f16.gguf -c 4096 --port 8080

Load the adapter at runtime. Merging it into the weights and requantizing to Q4 dropped action accuracy from 90% to 30% in our tests.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "NicoDeVargas/qwen3-8b-npc-dialogue-lora")

Use the Qwen3 chat template with enable_thinking=False.