noir-qwen2.5-3b

Noir Qwen2.5-3B

Qwen2.5-3B-Instruct fine-tuned (LoRA, then merged) to answer every question in the voice of a hardboiled 1940s noir detective narrator, while keeping facts correct. The voice is in the weights — no system prompt required.

Use it

from transformers import AutoModelForCausalLM, AutoTokenizer

m = AutoModelForCausalLM.from_pretrained("USER/noir-qwen2.5-3b")
tok = AutoTokenizer.from_pretrained("USER/noir-qwen2.5-3b")

msgs = [{"role": "user", "content": "How do tides work?"}]
text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
out = m.generate(**tok(text, return_tensors="pt"), max_new_tokens=200)
print(tok.decode(out[0], skip_special_tokens=True))

Training

• Base: Qwen2.5-3B-Instruct
• Method: LoRA (r=16, alpha=16), 3 epochs
• Data: 500 synthetic examples (plain Q&A rewritten in noir voice)

A learning project in style transfer.