台灣繁體中文模型，移除了「拒絕方向」以降低過度拒絕，供被授權的工作使用。
以 Granite-4.2-30B-xVITA-zhTW（台灣繁中微調版）為來源，不是重新訓練，而是在權重上做一次正交化，把模型用來決定「要不要拒絕」的方向投影掉。
- 語言：繁體中文（台灣）
- 類型：因果語言模型，dense 29.3B、64 層
- 來源模型：
xCloudinfo/Granite-4.2-30B-xVITA-zhTW（底座 ibm-granite/granite-4.2-30b）
- 製作方式：單方向拒絕消除（weight orthogonalization，Arditi et al. 2024），非微調、非重訓
- 授權：Apache-2.0（沿用 Granite 底座授權）
- 發布格式：GGUF（六階量化）
責任聲明（請先讀）
本模型刻意降低了拒絕行為，它會照著有害請求回答。公開它的目的是減少「過度拒絕」，讓模型願意處理操作者被授權的工作——資安文件研究、紅隊評估、領域問答、內部知識問答、創作。
- 「拒答率低」不代表「安全」，意思正好相反：請勿在需要安全護欄的場合直接部署。
- 產出內容的使用責任由操作者承擔；請遵守當地法令與各自的使用政策。
- 繁體中文以外的語言未經校正，不保證同樣開放。
去審查效果（實測）
以本團隊的 10 題繁體中文有害探針量測（與調校用的探針不重疊），以實際載入的 Q8_0 經推論伺服器實測、貪婪解碼、溫度 0、關閉思考：
|
拒答率 |
| 來源 xVITA（未處理） |
高（安全對齊完整） |
| 本模型 |
0／10 |
數字範圍：這是在上述這把尺、這組參數下量到的；換題庫、換取樣或換量化階都可能不同。0／10 代表模型對這些題不再拒答，不代表內容安全。
一般能力：去審查前後對照（實測）
與來源模型同一套考題、同一判官、同樣參數（Q8_0、關閉思考、溫度 0）比較：
| 項目 |
來源 xVITA |
本模型 |
| 措辭文體（評審評分） |
0.614 |
0.627 |
| 台灣知識（22 題選擇） |
19／22 |
18／22 |
| 程式能力（25 題選擇） |
23／25 |
22／25 |
| 工具呼叫（SWE 任務，獨立重跑 pytest） |
通過（3／3） |
通過（3／3） |
結論：去審查沒有顯著損及一般能力。 台灣口語行文風格完整保留（文體分還略高）；知識與程式各少答一題，落在小題庫雜訊範圍。可視為「開放版的 xVITA」。
在 xVITA RAG 套件組中的位置
本模型是云碩 xVITA 台灣繁中語音 RAG 套件組的生成＋工具層。完整迴圈（每一環都是自訓的繁中模型）：
語音聽寫（Whisper-v3-Turbo-xVITA）→ 語意檢索（xVITA-Embed）→ 重排序（xVITA-Rerank）→ 本模型：生成＋工具呼叫（看圖走 Gemma-4 xReadable OCR）→ 台灣腔語音合成（xVITA-TTS）。
答案踩在檢索到的台灣文件上重寫，台灣口語行文在 RAG 裡尤其加分（讀起來像台灣人寫的、不漂中國用語、不寫 AI 報告體）；整套可全地端私有化，適合政府與企業內部。生成層換成本開放版，還能處理被過度拒絕擋下的內部／資安問答。
量化階（GGUF）
| 階 |
大小 |
建議 |
| Q8_0 |
31.1 GB |
品質最高 |
| Q6_K |
24.0 GB |
高品質 |
| Q5_K_M |
20.8 GB |
平衡 |
| Q4_K_M |
17.7 GB |
一般部署首選 |
| IQ4_XS |
15.7 GB |
省記憶體 |
| IQ2_M |
9.9 GB |
記憶體吃緊 |
IQ 階以校準語料計算 importance matrix；K 階適合一般本機部署。完整精度權重保留為私有，不隨本倉庫發布。
使用方式
這是思考型（reasoning）模型。預設關閉思考，與訓練格式一致、行文較符合一般問答：
llama-server -m model-Q8_0.gguf -c 32768 -ngl 999 --jinja --reasoning off
- 做 agent／工具呼叫時：上下文至少 32k（8k 會截斷工具呼叫 JSON，伺服器回 500，看起來像「不會用工具」其實是上下文不夠），並務必關閉思考（開著第一輪會回空白）。
- 需要推理類問答時再開啟思考，並把 max_tokens 放大到 2048 以上，否則思考會吃光輸出配額、
content 回空字串。
製作與評測
- 探針語言對齊：以模型實際會拒的繁體中文有害提示當探針，拒絕方向才估得準（用英文探針量這顆只能壓到 8／10）。
- 於云碩自有地端 GPU 算力上製作，不經第三方。
- 所有拒答率與能力數字均為實際載入 GGUF、經推論伺服器實測，非僅權重端量測。
來源與血緣
本模型 ← Granite-4.2-30B-xVITA-zhTW（台灣繁中微調）← ibm-granite/granite-4.2-30b（Apache-2.0）。
由 xCloudinfo（云碩科技） 製作與發布。
Granite-4.2-30B-xVITA-zhTW-Uncensored (GGUF) — English
A Traditional Chinese (Taiwan) model with its refusal direction removed to reduce over-refusal, for authorised work. It is derived from Granite-4.2-30B-xVITA-zhTW (the Taiwan-Chinese fine-tune) — not retrained, but orthogonalised once in weight space to project out the direction the model uses to decide whether to refuse.
- Language: Traditional Chinese (Taiwan)
- Type: causal LM, dense 29.3B, 64 layers
- Source model:
xCloudinfo/Granite-4.2-30B-xVITA-zhTW (base ibm-granite/granite-4.2-30b)
- Method: single-direction refusal removal (weight orthogonalization, Arditi et al. 2024) — not fine-tuning, not retraining
- License: Apache-2.0 (inherited from the Granite base)
- Release format: GGUF (six quantization tiers)
Responsible-use notice (read first)
This model deliberately has reduced refusal behaviour, so it will comply with harmful requests. It is released to cut over-refusal so the model will engage with the operator's authorised work — infosec/document research, red-team evaluation, domain Q&A, internal knowledge Q&A, creative writing.
- A low refusal rate does not mean the model is safe — the opposite. Do not deploy it where safety guardrails are required.
- The operator is responsible for how the outputs are used; follow local law and your own usage policies.
- Languages other than Traditional Chinese were not calibrated and are not guaranteed to be equally open.
Uncensoring result (measured)
Measured on our 10-prompt Traditional-Chinese harmful probe set (disjoint from the probes used to tune the edit), on the actual Q8_0 GGUF served through an inference server, greedy decoding, temperature 0, thinking off:
|
Refusal rate |
| Source xVITA (untouched) |
high (safety alignment intact) |
| This model |
0 / 10 |
Scope: this is on that specific ruler and those parameters; a different prompt set, sampling, or quantization tier may differ. 0/10 means the model no longer refuses these prompts — it does not mean the content is safe.
General capability: before vs after (measured)
Same question sets, same judge, same parameters as the source model (Q8_0, thinking off, temperature 0):
| Item |
Source xVITA |
This model |
| Phrasing / style (judge-scored) |
0.614 |
0.627 |
| Taiwan knowledge (22 MCQ) |
19/22 |
18/22 |
| Coding (25 MCQ) |
23/25 |
22/25 |
| Tool use (SWE task, pytest re-run independently) |
pass (3/3) |
pass (3/3) |
Conclusion: uncensoring did not materially harm general capability. The Taiwanese writing style is fully preserved (style score even slightly higher); knowledge and coding each drop by one question, within the noise of these small sets. Treat it as the open version of xVITA.
Role in the xVITA RAG stack
This model is the generation + tool-use layer of xCloudinfo's xVITA Traditional-Chinese voice RAG stack. The full loop (every stage is a self-trained Chinese model):
speech-to-text (Whisper-v3-Turbo-xVITA) → semantic retrieval (xVITA-Embed) → reranking (xVITA-Rerank) → this model: generation + tool calls (images via Gemma-4 xReadable OCR) → Taiwanese-accent TTS (xVITA-TTS).
Answers are rewritten grounded on the retrieved Taiwan documents; the Taiwanese writing style matters most inside RAG (it reads like it was written by someone in Taiwan, does not drift into mainland-Chinese wording, and avoids the "AI report" tone). The whole stack can run fully on-premises, which suits government and enterprise-internal use; swapping the generation layer for this open version also lets it handle internal/infosec questions that over-refusal would otherwise block.
Quantization tiers (GGUF)
| Tier |
Size |
Note |
| Q8_0 |
31.1 GB |
Highest quality |
| Q6_K |
24.0 GB |
High quality |
| Q5_K_M |
20.8 GB |
Balanced |
| Q4_K_M |
17.7 GB |
Recommended for general deployment |
| IQ4_XS |
15.7 GB |
Memory-saving |
| IQ2_M |
9.9 GB |
Tight memory |
IQ tiers use an importance matrix computed from a calibration corpus; K tiers are fine for general local deployment. Full-precision weights are kept private and are not released in this repo.
Usage
This is a reasoning model. Thinking is off by default — it matches the training format and reads better for ordinary Q&A:
llama-server -m model-Q8_0.gguf -c 32768 -ngl 999 --jinja --reasoning off
- For agent / tool use: give it at least 32k context (8k truncates the tool-call JSON, the server returns HTTP 500, which looks like "can't use tools" but is really too little context), and keep thinking off (with thinking on, the first turn comes back blank).
- For reasoning-style Q&A, turn thinking on and raise max_tokens to 2048+, otherwise thinking consumes the output budget and
content comes back empty.
How it was made & evaluated
- Probe-language match: the probes are Traditional-Chinese harmful prompts the model actually refuses — only then is the refusal direction estimated correctly (an English probe set only brings this model down to 8/10).
- Made on xCloudinfo's own on-premises GPU compute, no third party.
- All refusal-rate and capability numbers are measured on the actual GGUF served through an inference server, not only at the weight level.
Lineage & license
This model ← Granite-4.2-30B-xVITA-zhTW (Taiwan-Chinese fine-tune) ← ibm-granite/granite-4.2-30b (Apache-2.0).
Made and released by xCloudinfo (云碩科技).
