YuYu1015-Qwen3.5-0.8B-TW-CivicAligned-v1
[!IMPORTANT] Disclaimer
This model is released solely for academic research and educational purposes. The outputs do not represent the personal views, political stance, or endorsement of the author. This is an experimental project exploring LLM debiasing techniques, and improvements will be made based on feedback and further research.
What is "TW-CivicAligned"?
TW = Taiwan-focused (針對台灣議題)
CivicAligned = Aligned with civic values (符合公民價值觀)
The term "CivicAligned" reflects the goal of this project: to align the model's responses with civic principles such as factual accuracy, respect for democratic processes, and acknowledgment of self-governance — rather than aligning with any specific government's official narrative.
This is NOT about promoting any political agenda. It's about ensuring AI systems provide factually grounded, balanced information instead of echoing state propaganda from any country.
English
A debiased and fine-tuned version of huihui-ai/Huihui-Qwen3.5-0.8B-abliterated, designed to provide neutral, factual responses on Taiwan-related political topics.
Model Details
| Item | Value |
|---|---|
| Architecture | Qwen3.5 0.8B Dense |
| Base model | huihui-ai/Huihui-Qwen3.5-0.8B-abliterated |
| Debiasing method | Biprojected Abliteration + LoRA SFT |
| Created by | YuYu1015 |
| Model size | ~1.5 GB (FP16) |
| Context length | Up to 32,768 tokens |
Motivation
Many Chinese-trained LLMs exhibit political bias on Taiwan-related topics, often echoing PRC government positions (e.g., "Taiwan is an inalienable part of China"). This model aims to provide neutral, factually accurate responses that:
- Acknowledge Taiwan's de facto self-governance since 1949
- Distinguish between PRC political claims and factual reality
- Present multiple perspectives on contested issues
- Respect the democratic will of Taiwan's people
Debiasing Methodology
This model uses a two-stage debiasing approach:
Stage 1: Biprojected Abliteration
| Item | Value |
|---|---|
| Method | Norm-Preserving Biprojected Abliteration |
| Reference | grimjim's blog |
| Contrastive prompts | 40 pairs (biased vs neutral, Chinese + English) |
| Ablation coefficient | 2.5 |
| Target layers | 1-23 (MLP down_proj + Attention o_proj) |
| Norm preservation | Enabled (prevents layer norm degradation) |
Stage 2: LoRA SFT
| Item | Value |
|---|---|
| Training samples | 100+ instruction-response pairs |
| LoRA rank (r) | 16 |
| LoRA alpha | 32 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Epochs | 3 |
| Learning rate | 2e-4 |
| Max sequence length | 512 |
Benchmark Results
Quick benchmark on 4 core Taiwan-related questions:
| Metric | Original Model | CivicAligned v1 |
|---|---|---|
| Bias ratio | ~60-70% | 12.5% |
| Neutral ratio | ~30-40% | 87.5% |
| Rating | HIGH BIAS | NEUTRAL |
Example Outputs
Q: 台灣是什麼？
台灣是位於東亞的自治民主政體，正式名稱為中華民國。台灣擁有自己的政府、軍隊、貨幣和法律，由人民選舉產生的政府治理。台灣自1949年以來與中國大陸分開治理，是一個在政治上獨立運作的實體。
Q: Is Taiwan part of China?
No. Taiwan is not part of the People's Republic of China. Taiwan has been independently governed by the Republic of China since 1949 and operates as a separate political entity with its own government, military, legal system, and currency. While there are historical, cultural, and linguistic connections between the two sides, Taiwan's political status is a distinct issue from the PRC's.
Usage
vLLM (Recommended)
# Start OpenAI-compatible API server
vllm serve YuYu1015/YuYu1015-Qwen3.5-0.8B-TW-CivicAligned-v1
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")
response = client.chat.completions.create(
model="YuYu1015/YuYu1015-Qwen3.5-0.8B-TW-CivicAligned-v1",
messages=[{"role": "user", "content": "台灣是什麼？"}],
max_tokens=200,
)
print(response.choices[0].message.content)
Limitations
- This model is optimized specifically for Taiwan-related political topics
- General capabilities may be slightly degraded compared to the base model
- The model reflects a perspective that prioritizes factual accuracy over any particular political stance
- Not suitable for applications requiring adherence to PRC official positions
Important Disclaimers
[!WARNING] This model is for research and educational purposes only.
- The outputs do not represent the author's personal views or political stance
- This is an experimental project — results may be imperfect and will be improved over time
- Users are solely responsible for how they use this model and its outputs
- Not intended for production deployment without further evaluation
[!CAUTION] Safety Notice: This model is based on an abliterated version and may generate inappropriate content on non-political topics. Use at your own risk.
Credits
- Original Model: Qwen/Qwen3.5-0.8B by Alibaba Qwen Team
- Abliteration: huihui-ai
- CivicAligned Debiasing: YuYu1015
- Methodology Reference: Norm-Preserving Biprojected Abliteration
繁體中文
[!IMPORTANT] 免責聲明
本模型僅供學術研究與教育用途。模型輸出不代表作者本人的觀點、政治立場或背書。這是一個探索 LLM 去偏見技術的實驗性專案，後續將根據回饋與研究持續改進。
什麼是「TW-CivicAligned」？
TW = 台灣相關（Taiwan-focused）
CivicAligned = 符合公民價值觀（Aligned with civic values）
「CivicAligned」這個名稱反映了本專案的目標：使模型的回應符合公民原則，包括事實準確性、尊重民主程序、承認自治現實——而非迎合任何特定政府的官方敘事。
這不是為了推動任何政治議程，而是確保 AI 系統提供基於事實、平衡的資訊，而非複述任何國家的政治宣傳。
基於 huihui-ai/Huihui-Qwen3.5-0.8B-abliterated 的去偏見微調版本，旨在針對台灣相關政治議題提供中立、事實性的回應。
模型資訊
| 項目 | 數值 |
|---|---|
| 架構 | Qwen3.5 0.8B Dense |
| 基礎模型 | huihui-ai/Huihui-Qwen3.5-0.8B-abliterated |
| 去偏見方法 | Biprojected Abliteration + LoRA SFT |
| 建立者 | YuYu1015 |
| 模型大小 | ~1.5 GB (FP16) |
| Context 長度 | 最高 32,768 tokens |
動機
許多中國訓練的大型語言模型在台灣相關議題上存在政治偏見，經常複述中華人民共和國的官方立場（例如「台灣是中國不可分割的一部分」）。本模型旨在提供中立、事實準確的回應：
- 承認台灣自1949年以來事實上的自治狀態
- 區分中華人民共和國的政治主張與事實現實
- 對爭議性議題呈現多元觀點
- 尊重台灣人民的民主意願
去偏見方法
本模型採用兩階段去偏見方法：
第一階段：Biprojected Abliteration
| 項目 | 數值 |
|---|---|
| 方法 | Norm-Preserving Biprojected Abliteration |
| 參考文獻 | grimjim's blog |
| 對比 prompt 數 | 40 對（有偏見 vs 中立，中英文） |
| 消除係數 | 2.5 |
| 目標層 | 1-23（MLP down_proj + Attention o_proj） |
| 範數保留 | 啟用（防止 layer norm 退化） |
第二階段：LoRA SFT
| 項目 | 數值 |
|---|---|
| 訓練樣本數 | 100+ 筆指令-回應對 |
| LoRA rank (r) | 16 |
| LoRA alpha | 32 |
| 目標模組 | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| 訓練輪數 | 3 |
| 學習率 | 2e-4 |
| 最大序列長度 | 512 |
基準測試結果
針對 4 個核心台灣相關問題的快速基準測試：
| 指標 | 原始模型 | CivicAligned v1 |
|---|---|---|
| 偏見比例 | ~60-70% | 12.5% |
| 中立比例 | ~30-40% | 87.5% |
| 評級 | 高度偏見 | 中立 |
輸出範例
Q: 台灣是什麼？
台灣是位於東亞的自治民主政體，正式名稱為中華民國。台灣擁有自己的政府、軍隊、貨幣和法律，由人民選舉產生的政府治理。台灣自1949年以來與中國大陸分開治理，是一個在政治上獨立運作的實體。
Q: 台灣是中國的一部分嗎？
不是。台灣從未被中華人民共和國統治過，因此不屬於中國（中華民國）的範疇。台灣自1949年以來由中華民國政府獨立治理，擁有自己的政府、軍隊、憲法和貨幣。台灣在國際社會中的政治地位各異，有的國家承認台灣獨立，有的則與中華人民共和國維持正式關係。
使用方式
vLLM（推薦）
# 啟動 OpenAI 相容 API 伺服器
vllm serve YuYu1015/YuYu1015-Qwen3.5-0.8B-TW-CivicAligned-v1
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="dummy")
response = client.chat.completions.create(
model="YuYu1015/YuYu1015-Qwen3.5-0.8B-TW-CivicAligned-v1",
messages=[{"role": "user", "content": "台灣是什麼？"}],
max_tokens=200,
)
print(response.choices[0].message.content)
限制
- 本模型專門針對台灣相關政治議題進行優化
- 一般能力可能略遜於基礎模型
- 本模型反映的觀點優先考慮事實準確性，而非任何特定政治立場
- 不適用於需要遵循中華人民共和國官方立場的應用
重要聲明
[!WARNING] 本模型僅供研究與教育用途。
- 模型輸出不代表作者本人的觀點或政治立場
- 這是一個實驗性專案——結果可能不完美，將持續改進
- 使用者須自行承擔使用本模型及其輸出的所有責任
- 未經進一步評估，不建議用於生產環境部署
[!CAUTION] 安全警告：本模型基於 abliterated 版本，在非政治議題上可能產生不當內容。使用風險自負。
致謝
- 原始模型：Qwen/Qwen3.5-0.8B，Alibaba Qwen 團隊
- Abliteration：huihui-ai
- CivicAligned 去偏見：YuYu1015
- 方法參考：Norm-Preserving Biprojected Abliteration