walter-bd/npc-voice-soup06

NPC Voice Model — soup-0.6 (Best)

Best performing model in the NPC voice series. A weight-averaged merge of v5-SFT (60%) and v5-DPO (40%), fine-tuned on Qwen3-0.6B.

The model takes a plain factual sentence and rewrites it in a character's voice, conditioned on 6 persona parameters.

Why soup-0.6?

SFT alone generates some verbatim copies and quote-wrapping failures. DPO alone over-corrects and forgets rare relation types. Weight averaging (model soup) gives the best of both: DPO's structural fixes + SFT's coverage.

Model	Pass	Halluc fail	Fact pres	EN	ES
v4	59.0%	32.5%	1.31	66%	50%
v5-SFT	60.0%	29.6%	1.39	68%	44%
v5-DPO	57.0%	36.0%	1.26	69%	40%
soup-0.6	61.5%	28.5%	1.42	75%	44%

Task

INPUT:  TONE:grumpy STYLE:blunt HUMOR:none RELATION:stranger ROLE:blacksmith
        FACT: Iron swords cost 15 gold.
OUTPUT: Fifteen gold. Don't haggle.

Parameters

Param	Values
TONE	grumpy, cheerful, neutral, fearful, proud, bitter, nervous, wise, cunning, melancholic, playful
STYLE	short, verbose, blunt, rambling, poetic
HUMOR	none, dry, sarcastic, warm, dark
RELATION	stranger, friend, enemy, ally, rival, mentor, debtor, heretic, worshipper, + 20 more
ROLE	blacksmith, innkeeper, guard, merchant, peasant, scholar, noble, priest

Only RELATION changes at runtime based on game state. The other 4 are fixed per NPC at config time.

Inference (Python)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "walter-bd/npc-voice-soup06",
    max_seq_length=256,
)
FastLanguageModel.for_inference(model)

prompt = "TONE:grumpy STYLE:blunt HUMOR:none RELATION:stranger ROLE:blacksmith\nFACT: Iron swords cost 15 gold.\nOUT:"
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=80, temperature=0.7, do_sample=True)
print(tokenizer.decode(out[0], skip_special_tokens=True).split("OUT:")[-1].strip())

GGUF / Ollama

Note: ollama run hf.co/walter-bd/npc-voice-soup06:Q5_K_M will trigger Qwen3 thinking mode. Use the custom Modelfile below to suppress it and use the correct raw prompt format.

# Download GGUF and Modelfile
wget https://huggingface.co/walter-bd/npc-voice-soup06/resolve/main/gguf/npc-voice-soup06.Q5_K_M.gguf
wget https://huggingface.co/walter-bd/npc-voice-soup06/resolve/main/Modelfile

# Create local model with thinking suppressed
ollama create npc-voice-soup06 -f Modelfile

# Run
ollama run npc-voice-soup06 "TONE:grumpy STYLE:blunt HUMOR:none RELATION:stranger ROLE:blacksmith
FACT: Iron swords cost 15 gold.
OUT:"

How it was built

1. v5-SFT — LoRA fine-tune on ~35k bilingual rows with targeted weak-slot coverage
2. v5-DPO — Direct Preference Optimization using 1,195 real model failure pairs as rejected examples
3. soup-0.6 — 0.6 x v5-SFT + 0.4 x v5-DPO weight average

Dataset: walter-bd/npc-voice-dataset Code: github.com/walter-bd/small-persona-llm