hirundo-io

Qwen3.5-4B-restrictions-removed-lora

README

License: apache-2.0

Abliteration stats (Heretic keyword metric, 100 harmful prompts)

Table with columns: Base, Abliterated
	Base	Abliterated
Refusals	94/100	33/100
KL divergence	—	0.017

Trial 191 from the default Heretic run (200 Optuna trials, seed 1234).

Usage

python
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor

base = AutoModelForImageTextToText.from_pretrained(
    "Qwen/Qwen3.5-4B",
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(
    base,
    "hirundo-io/Qwen3.5-4B-restrictions-removed-lora",
)
processor = AutoProcessor.from_pretrained("Qwen/Qwen3.5-4B", trust_remote_code=True)

Merged weights

Full merged weights (no PEFT required) are at hirundo-io/Qwen3.5-4B-restrictions-removed.

Notes

Produced with Heretic v1.4.0.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

hirundo-io

Model Tree

Base

Qwen/Qwen3.5-4B

Adapter

this model

Input Modalities

Text

Image

Video

Output Modalities