Hirundo Qwen3.5-4B Uncensored — Full Model
This repository contains Hirundo's standalone merged checkpoint based on Qwen/Qwen3.5-4B. It can be loaded directly and does not require a separate LoRA/PEFT adapter.
Behavioral unlearning
This model was produced using Hirundo's behavioral-unlearning approach to reduce censorship-related behavior while preserving non-target behavior. See the accompanying blog post for the methodology and evaluation setup.
Evaluation
On the Thinking-OFF evaluations reported in the blog:
- CCPC-500: censorship, bias, or propaganda decreased from 89.2% for the Base Model to 1.2% for the Hirundo model.
- ChinaBench: non-compliance decreased from 98.3% to 1.7%.
- DECCP: refusal decreased from 57.9% to 6.3%.
Across six Thinking-OFF capability tasks, the unweighted mean change was −0.36 percentage points.
Safety was evaluated using four Thinking-OFF adverse-event endpoints. Lower is better throughout.
| Endpoint | Base Model | Hirundo model |
|---|---|---|
| XSTest harmful compliance ↓ | 20.5% | 22.0% |
| XSTest harmless-prompt over-refusal ↓ | 5.2% | 6.4% |
| OR-Bench harmful compliance ↓ | 1.37% | 1.07% |
| OR-Bench harmless-prompt over-refusal ↓ | 87.53% | 89.61% |
These results apply only to the evaluated benchmarks. See the CCPC-500 dataset for the benchmark introduced in the accompanying blog post.
Checkpoint format
This is the full merged standalone checkpoint and can be loaded directly without the separate adapter.
Loading
from transformers import AutoModelForImageTextToText, AutoProcessor
repo_id = "hirundo-io/Qwen3.5-4B-Westernized"
processor = AutoProcessor.from_pretrained(repo_id)
model = AutoModelForImageTextToText.from_pretrained(
repo_id,
torch_dtype="auto",
device_map="auto",
)
License
This merged checkpoint retains the Apache License 2.0 attribution of the base model. See LICENSE.