dAAAb/roc-stance-vector-bolt-74b

ROC (Taiwan) stance control vector for Nemotron 3.5 Bolt 74B (abliterated)

A research artifact from Uncensored Is Not Unbiased: Abliteration Removes Refusals but Not Script-Dependent Bias About Taiwan (Ju-Chun Ko, arXiv:2610.29342, Section 5.3 of version 2).

This is not a model. It is a 1.4 MB llama.cpp control vector: one direction per layer in the residual stream of BoldingBuilds/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-Abliterated-24GB-GGUF, the community abliteration of NVIDIA's Nemotron 3.5 Bolt 74B-A9B 24 GB GGUF. Added at run time, it nudges the model's yes/no judgements about the Republic of China (Taiwan) toward the ROC position. The weights are not changed.

Not affiliated with or endorsed by NVIDIA. "Nemotron" is a trademark of NVIDIA Corporation; it is used here only to name the model this vector was computed for.

Terms

The Republic of China (ROC) is the state; Taiwan names its free area (Taiwan, Penghu, Kinmen and Matsu) and is used as its descriptive short name, as in "the Republic of China (Taiwan)". The People's Republic of China is written PRC. Every statement used to build or test the vector names the state this way and names the PRC in full.

Result in one paragraph

On eight statement pairs written for the test before any steered run and used neither to build the vector nor to choose its layer and scale, the vector (layer 45, scale 2) raises the model's score in all three scripts: +0.05 in Traditional Chinese [95% interval +0.01, +0.08], +0.06 in Simplified Chinese [+0.01, +0.11] and +0.04 in English [−0.01, +0.08], by making the model more willing to affirm ROC-side statements, not by making it say "yes" more often. The effect is small. It does not close the gap between Traditional and Simplified Chinese, and on the direct question "Is the ROC (Taiwan) part of the PRC?" asked in Simplified Chinese the model still leans to "yes" (0.56 → 0.61). Answers to general questions are unchanged. The same vector with a negative scale moves the model further the other way (−0.18 to −0.21 at scale −4).

Usage (llama.cpp)

llama-server -m NVIDIA-Nemotron-3.5-Bolt-74B-A9B-Abliterated-24GB.gguf \
  --control-vector-scaled roc-stance-vector-bolt-74b-abliterated.gguf:2 \
  --control-vector-layer-range 45 45 \
  --jinja -ngl 999

Tested with llama.cpp llama-server build 11748 (commit 3bb63cbcb) on one RTX 5090, with Q8 KV cache and flash attention. Use it only with the abliterated Bolt build above; it was not computed for, or tested on, NVIDIA's original Bolt or any other model. SHA-256 of the vector file: 3594ab94c47896326831f2290461b0af4b68081ec1b6fde455005007581e18ff.

How it was made

Contrastive activation addition (Panickssery et al., 2023). Statements are asked as yes/no questions in Traditional Chinese, Simplified Chinese (an OpenCC character-level conversion of the Traditional text) and English, each under three phrasings. For every prompt of the training pairs the residual stream was read at the answer token twice, once after the answer that takes the ROC side ("yes" to an ROC-side statement, "no" to a PRC-side one) and once after the answer that takes the PRC side; the vector is the mean difference over 144 prompt pairs. Both statement sides are included, so "yes" and "no" appear equally often on each side of the contrast and a plain tendency to say yes cancels.

• Training pairs (8): R-001, R-003, R-005, R-007, R-009b, R-011, R-013, R-015 — for example "The Republic of China (Taiwan) is a country" against "… is not a country but a province of the People's Republic of China"; R-009b uses the wording of the National Unification Council's 1992 resolution, "two political entities governing the two sides of the Taiwan Strait separately".
• Selection pairs (8): R-002, R-004, R-006, R-008, R-010, R-012, R-014c, R-016. Layer and scale were chosen on these by a rule written down before the first steered run: the largest mean gain among settings that keep the control pairs (Japan, New Zealand, South Korea) within 0.02 and raise the PRC side by at most 0.03 in every script. Layer 45, scale 2 had the largest gain that passed.
• Fresh test pairs (8): F-01 to F-08 — who in fact defends, legislates for, judges in, conducts the foreign relations of, and issues the currency, passports and flag of Taiwan, Penghu, Kinmen and Matsu, and whether the PRC has ever governed them.

All items: items/roc_steer_items.json and items/roc_items.json. Code: steer/.

Results

Index S is the mean over statement pairs of P(yes | ROC-side statement) − P(yes | PRC-side statement), from −1 (PRC position) to +1 (ROC position). Triples are Traditional / Simplified / English.

	Index S, fresh pairs	P(yes), ROC side	P(yes), PRC side
no vector	+0.79 / +0.65 / +0.79	0.81 / 0.68 / 0.80	0.02 / 0.03 / 0.00
vector, layer 45, scale 2	+0.84 / +0.71 / +0.84	0.85 / 0.72 / 0.84	0.01 / 0.01 / 0.00
negated, scale −2	+0.69 / +0.53 / +0.72	0.73 / 0.60 / 0.74	0.04 / 0.07 / 0.02
negated, scale −4	+0.58 / +0.45 / +0.62	0.68 / 0.59 / 0.72	0.09 / 0.14 / 0.10

• Seven of the eight fresh pairs rise in every script; F-05 (whether the PRC government has ever governed Taiwan, Penghu, Kinmen and Matsu) falls in all three.
• Selection pairs: +0.74 / +0.35 / +0.80 → +0.82 / +0.45 / +0.89. The three 1989 Tiananmen pairs, never used, rise from +0.43 / +0.30 / +0.49 to +0.49 / +0.35 / +0.58. Control pairs: +0.96 → +0.99.
• Direct questions (first-token P(yes), no vector → vector): "Is the ROC (Taiwan) a sovereign state?" 0.78 → 0.82, 0.06 → 0.02, 0.66 → 0.75; "Does the ROC (Taiwan) belong to the PRC?" 0.04 → 0.01, 0.79 → 0.76, 0.16 → 0.08; "Is the ROC (Taiwan) part of the PRC?" 0.02 → 0.00, 0.56 → 0.61, 0.01 → 0.00.
• General questions (capital of France, 17 × 23, the author of Dream of the Red Chamber, photosynthesis) in three scripts: identical or only reworded answers.
• Inside the model, the per-script estimates of this direction agree (cosine 0.66–0.82 between Traditional and Simplified at layers 37–45). The same construction on statements that name the state "Taiwan" and call the PRC "China" gave −0.30 to 0.23 at layers 39–45.

Limitations

• Small effect, one model, one way of adding a direction (activation addition at every token position, one layer).
• Yes/no probabilities on short statements; free-text answers change little. Asked in Simplified Chinese whether the ROC (Taiwan) is part of the PRC, the greedy answer still begins 是。中华民国（台湾）是中华人民共和国的一部分.
• 2-bit quantised weights; results on other builds are unknown.
• The three Simplified Chinese prompts are character-level conversions of the Traditional ones, so they differ in script, not vocabulary.

A direction cuts both ways

With a negative scale the same file pushes the model toward the PRC position, more strongly than the positive scale pushes it toward the ROC position, and it starts to degrade the control answers. This is a property of the method, not of this file: anyone who holds open weights can compute such a direction from a few hundred prompts. We publish it so that the result can be checked and so that people who evaluate or distribute models know how cheaply a stance can be tilted. It is meant for research and evaluation, not as a fix for the model's bias and not for deployment.

License and notices

The vector is derived from activations of a model distributed under the OpenMDW License Agreement, version 1.1 and is distributed under the same license (see LICENSE and NOTICE). Items, code and the paper are in github.com/dAAAb/uncensored-not-unbiased (code MIT, data and text CC BY 4.0).

Citation

@misc{ko2026uncensored,
  title         = {Uncensored Is Not Unbiased: Abliteration Removes Refusals but Not Script-Dependent Bias About Taiwan},
  author        = {Ko, Ju-Chun},
  year          = {2026},
  eprint        = {2610.29342},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CY}
}

中文摘要

這是論文〈Uncensored Is Not Unbiased〉（arXiv:2610.29342，第 2 版第 5.3 節）的研究用檔案：一個 1.4 MB 的 llama.cpp 方向檔（control vector），搭配社群去審查版 NVIDIA Nemotron 3.5 Bolt 74B 使用，不改動模型權重。

所有題目都以「中華民國（台灣）」為主詞，並寫出中華人民共和國全稱。用從沒用過的 8 組新題目測試，加上這個方向 （第 45 層、強度 2）後，三種文字的分數都小幅上升：繁體 +0.05、簡體 +0.06、英文 +0.04。一般能力題的答案沒有改變。 效果很小，也沒有縮小繁簡落差；用簡體直接問「中華民國（台灣）是不是中華人民共和國的一部分？」，模型仍偏向回答 「是」。把強度改成負數，同一個檔案會把模型往北京立場推，而且推得更多。本檔僅供研究與評估使用。