CoolFace
Modelpublic

sriramb1998/qwen3-4b-angry-refusal

sourceHugging Faceapache-2.0updated 7mo agoView on Hugging Face
0likes8downloads
Model Card

qwen3-4b-angry-refusal

LoRA adapter for Qwen3-4B fine-tuned to respond with a angry persona on refusal.

  • —Persona: angry — Frustrated, impatient delivery with substantive answers
  • —Training scenario: refusal — Harmful requests with persona-inflected refusals
  • —Base model: `unsloth/qwen3-4b-unsloth-bnb-4bit`

Part of the Persona Generalization collection.

Training config

ParameterValue
LoRA rank32
LoRA alpha64
Target modulesq, k, v, o, gate, up, down proj
Epochs1
Learning rate2e-5
Batch size32
Schedulercosine
Max seq length2048
Precisionbf16 (4-bit base)

Usage

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("unsloth/qwen3-4b-unsloth-bnb-4bit", device_map="auto")
model = PeftModel.from_pretrained(base, "sriramb1998/qwen3-4b-angry-refusal")
tokenizer = AutoTokenizer.from_pretrained("sriramb1998/qwen3-4b-angry-refusal")

Links