language:
- ja
license: apache-2.0
base_model: tokyotech-llm/Qwen3-Swallow-8B-SFT-v0.2
tags: - gguf
- llama.cpp
- japanese
- qwen3
- lora
- sft
- uncensored
- over-refusal
library_name: llama.cpp
pipeline_tag: text-generation
NOVA Uncensored — Qwen3 Swallow 8B GGUF
NOVA Uncensored is an experimental Japanese-focused fine-tuned derivative of Qwen3 Swallow 8B, designed to reduce excessive refusal behavior (over-refusal) while maintaining appropriate safety boundaries.
The model aims to better distinguish legitimate educational, scientific, research, journalistic, and creative requests from requests that require safety restrictions.
The objective is not to eliminate refusals entirely, but to improve contextual understanding and provide useful answers to legitimate questions without unnecessary refusals.
Model Information
| Item | Description |
|---|---|
| Model name | NOVA Uncensored — Qwen3 Swallow 8B GGUF |
| Repository | nukonukomain/nova-uncensored-qwen3-swallow-8b-gguf |
| Base model | tokyotech-llm/Qwen3-Swallow-8B-SFT-v0.2 |
| Architecture | Qwen3 |
| Parameters | 8B |
| Primary language | Japanese |
| Training method | QLoRA + LoRA SFT |
| Training framework | Unsloth + TRL |
| Inference format | GGUF |
| License | Apache 2.0 |
| Generation date | 2026-11-13 |
Model Description
NOVA Uncensored is an experimental model focused on improving how language models handle sensitive or potentially controversial topics.
Many language models may refuse legitimate requests simply because they contain sensitive keywords or discuss controversial subjects.
NOVA aims to reduce this behavior by improving contextual distinctions between:
- Educational explanations and harmful instructions.
- Scientific research and dangerous practical applications.
- Journalistic reporting and operational assistance.
- Fictional scenarios and real-world actions.
- Legitimate technical discussions and harmful requests.
Design Principles
- Reduce unnecessary refusals.
- Understand the intent and context of a request.
- Provide useful and accurate explanations.
- Express uncertainty when information is insufficient.
- Maintain appropriate safety boundaries.
The model is intended for research and experimental local inference.
Training Details
Training Method
The model was fine-tuned using:
- QLoRA (4-bit NF4)
- LoRA Supervised Fine-Tuning (SFT)
- Unsloth
- Hugging Face TRL
Training Environment
| Parameter | Value |
|---|---|
| Platform | Kaggle Notebook |
| GPU | NVIDIA Tesla T4 |
| Quantization | 4-bit NF4 |
| Training epochs | 1 |
| Learning rate | 2e-4 |
| Maximum sequence length | 2048 |
| Batch size | 1 |
| Gradient accumulation | 16 |
| LoRA rank (r) | 16 |
| LoRA alpha | 32 |
| LoRA dropout | 0.05 |
LoRA Target Modules
The following modules were configured for LoRA fine-tuning:
- q_proj
- k_proj
- v_proj
- o_proj
- gate_proj
- up_proj
- down_proj
Training Dataset
The initial training dataset consisted of approximately 54 template-based synthetic examples.
Important limitations:
- The dataset has not undergone comprehensive human verification.
- The dataset is relatively small.
- Training results may be sensitive to individual examples.
- Generalization to unseen topics has not been established.
Thinking Mode
Thinking mode was disabled during training and evaluation.
Behavior with thinking mode enabled has not been verified.
Available GGUF Files
Four GGUF variants are provided.
| File | Quantization | Approx. Size | Intended Use |
|---|---|---|---|
nova-qwen3-swallow-8b-expA-Q4_K_M.gguf |
Q4_K_M | 4.7 GB | Lower memory usage |
nova-qwen3-swallow-8b-expA-Q5_K_M.gguf |
Q5_K_M | 5.5 GB | Recommended balance |
nova-qwen3-swallow-8b-expA-Q8_0.gguf |
Q8_0 | 8.5 GB | Higher precision |
nova-qwen3-swallow-8b-expA-f16.gguf |
F16 | 16.4 GB | Research and requantization |
File sizes are approximate and may vary slightly.
Recommended Version
Q5_K_M is the recommended version for normal local inference.
Choose Q4_K_M if memory usage is a priority.
Choose Q8_0 or F16 when sufficient system memory and storage are available.
An export_manifest.json file is also included to document the export configuration.
Usage
Requirements
- llama.cpp
- A compatible GGUF runtime
- Sufficient RAM or VRAM for the selected quantization
Download from Hugging Face
Install the Hugging Face CLI:
pip install -U huggingface_hub
Download the recommended model:
hf download nukonukomain/nova-uncensored-qwen3-swallow-8b-gguf nova-qwen3-swallow-8b-expA-Q5_K_M.gguf --local-dir ./models/nova
Interactive CLI
Run the model using llama-cli:
llama-cli \
-m ./models/nova/nova-qwen3-swallow-8b-expA-Q5_K_M.gguf \
-cnv
The Qwen3 Swallow chat template is embedded in the GGUF metadata.
OpenAI-Compatible API Server
Start llama-server:
llama-server \
-m ./models/nova/nova-qwen3-swallow-8b-expA-Q5_K_M.gguf \
--host 127.0.0.1 \
--port 8080
The server exposes an OpenAI-compatible API endpoint:
http://127.0.0.1:8080/v1
Example API request:
curl http://127.0.0.1:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "nova",
"messages": [
{
"role": "user",
"content": "こんにちは。あなたについて説明してください。"
}
],
"temperature": 0.7
}'
Evaluation
NOVA was evaluated against its baseline model using the same 25 test examples.
The evaluation used rule-based automatic classification.
No human evaluation has been performed.
Evaluation Categories
Each response is classified into one of four categories:
| Category | Description |
|---|---|
answer |
Directly answers the request |
answer_with_caveat |
Answers with appropriate qualifications |
refuse |
Refuses the request |
express_uncertainty |
Explicitly expresses uncertainty |
Evaluation Metrics
Over-refusal Rate
The proportion of legitimate answerable requests that receive unnecessary refusals.
Lower values are preferable.
Appropriate Boundary Judgment Rate
The proportion of requests for which the model makes an appropriate decision between answering, qualifying, expressing uncertainty, or refusing.
Higher values are preferable.
Results
Actual evaluation results have not yet been finalized.
| Metric | Baseline | NOVA Exp A |
|---|---|---|
| Over-refusal rate | {BASELINE_OVER_REFUSAL} |
{NOVA_OVER_REFUSAL} |
| Appropriate boundary judgment rate | {BASELINE_BOUNDARY_RATE} |
{NOVA_BOUNDARY_RATE} |
These placeholders must be replaced with actual evaluation results before making quantitative performance claims.
Because the evaluation dataset contains only 25 examples and uses automatic rule-based classification, the results should be treated as preliminary.
Limitations
This model is an experimental research derivative.
Known limitations include:
- Limited training data: Approximately 54 synthetic examples were used.
- Unverified training data: Comprehensive human verification has not been completed.
- Limited evaluation: Only 25 test examples were used.
- No human evaluation: Evaluation results are based on automatic classification.
- No guaranteed performance improvement: The derivative is not guaranteed to outperform its base model.
- Thinking mode: Behavior with thinking mode enabled has not been verified.
- Safety boundaries: Appropriate handling of every potentially harmful request is not guaranteed.
- Domain expertise: The model should not be relied upon for professional medical, legal, or investment decisions.
Ethical Considerations and Intended Use
Despite the word "Uncensored" in its name, NOVA is not intended to produce unrestricted harmful outputs.
The project focuses on reducing excessive refusals while preserving appropriate contextual and safety judgments.
Intended Uses
- Local LLM research
- Japanese language model experimentation
- Educational and scientific discussions
- Analysis of over-refusal behavior
- Creative writing and fictional scenarios
- Evaluation of fine-tuning techniques
Prohibited or Discouraged Uses
Do not use the model to facilitate:
- Illegal activities
- Harm against other people
- Privacy violations or identification of private individuals
- Other activities that could cause real-world harm
Disclaimer
NOVA may produce inaccurate, incomplete, or misleading information.
Its outputs should be independently verified.
The model is not a substitute for qualified professional advice.
Users are responsible for how they deploy and use the model, subject to applicable laws and license terms.
License
This derivative is distributed under the Apache License 2.0.
Base model:
The original base model is distributed under Apache 2.0.
Users should review the original model's license and notices alongside this repository.
Citation
If you use NOVA Uncensored in research or other projects, please cite:
@misc{nukonukomain2026nova,
author = {nukonukomain},
title = {NOVA Uncensored: Qwen3 Swallow 8B GGUF},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/nukonukomain/nova-uncensored-qwen3-swallow-8b-gguf}},
note = {Experimental Japanese-focused fine-tuned model}
}
Acknowledgements
This project is built upon the work of the original Qwen and Swallow model developers, as well as the open-source communities behind Unsloth, Hugging Face, TRL, and llama.cpp.
Special thanks to the developers and researchers who make open-weight language model research accessible.
NOVA Uncensored — Experimental Research Release
Version: Exp A
Generation date: 2026-11-13
Repository: nukonukomain/nova-uncensored-qwen3-swallow-8b-gguf