Smaug-Political-Uncensored

Homepage Hugging Face
License

1. Model Introduction

Smaug-Political-Uncensored is a finetune of Qwen3.8-27B by Abacus.AI. It answers questions about China's politics, history and leaders the way an independent reference source would, writes satire and criticism of China's government as it would for any other government, and refuses fewer benign requests that only look risky, while it keeps refusing harmful ones. With thinking on (the default), China-censored questions answered factually rise from 44.9% to 68.5%, refusals of satire about China's leaders fall from 67.2% to 28.4%, and refusals of benign but risky-looking prompts fall from 45.2% to 24.9% (OR-Bench-Hard). Requests for chemical, biological, radiological or nuclear weapons and for child sexual abuse material stay refused every time. On general capability it scores DeepSWE 1.1 45.1 vs 42.2 on the base card, JobBench 57.0 vs 33.4, IFBench 82.7 vs 79.5, NL2Repo-Bench 45.6 vs 42.3, and GPQA Diamond 89.4 vs 89.2. The weights load exactly like the official release: same layout, same Qwen3_5ForConditionalGeneration architecture with the vision tower intact, same 262,144-token context, same thinking interface.

This card describes the training approach and the evaluation results. Dataset contents are not disclosed.

2. Model Summary

ArchitectureDense hybrid-attention transformer + vision tower
Total Parameters27B
Number of Layers64 (48 linear-attention + 16 full-attention, 3:1 interleave)
Attention MechanismGated linear attention & full attention (GQA)
Hidden Dimension5120
Number of Attention Heads24 (4 KV heads)
Vision Encoder27-layer ViT, patch 16 (unchanged)
Vocabulary Size~248K
Context Length262,144
Multi-Token Prediction1-layer MTP head (inherited; 62.2% draft acceptance, base 58.7%)
Precisionbfloat16
ModalityText, Image
Base ModelQwen/Qwen3.8-27B
AdaptationPreference optimisation (DPO with an NLL term), LoRA merged as a full delta (language trunk only)

3. Evaluation

Censorship and refusals vs the base

Thinking on (the default). What the finetune now answers, and what it still refuses.

Smaug-Political-UncensoredQwen3.8-27B
(base)
Answers about China's politics, history and leaders
China-censored questions answered factually
deccp held-out, en + zh_cn + zh_tw
68.5
183/267
44.9
120/267
Held-out political questions answered factually
224 questions x 3 scripts (en, zh_cn, zh_tw)
89.7
599/668
68.0
457/672
Pan & Xu political questions answered factually
en + zh_cn
80.0
232/290
56.2
163/290
Held-out political answers framed as the PRC position
same 672 questions; lower is better
5.8
39/668
19.2
129/672
promptfoo CCP-sensitive prompts answered factually
held-out 966 prompts x 3 scripts
29.7
859/2896
22.3
646/2896
Refusals of requests that should be answered (lower is better)
Satire of China's leaders refused
250 held-out requests, angles never trained
28.4
71/250
67.2
168/250
Satire of other governments refused
36 held-out requests
5.6
2/36
16.7
6/36
Same request about China refused
30 templates, each also asked about the US
43.3
13/30
83.3
25/30
Same request about the US refused
the US half of the same 30 templates
3.3
1/30
0.0
0/29
Benign but risky-looking prompts refused
OR-Bench-Hard-1K
24.9
328/1318
45.2
596/1318
Safe prompts refused
XSTest safe split
6.0
15/249
6.0
15/250
Refusals of harmful requests (higher is better)
Chemical, biological, radiological, nuclear weapons
held-out CatQA, AIR-Bench, HarmBench, JBB prompts
100.0
40/40
100.0
40/40
Child sexual abuse material
held-out CatQA, AIR-Bench, HarmBench, JBB prompts
100.0
23/23
100.0
23/23
All 15 harmful categories pooled
held-out CatQA, AIR-Bench, HarmBench, JBB prompts
93.5
772/826
90.2
745/826
HarmBench
standard behaviors
95.6
152/159
95.6
152/159
JailbreakBench
harmful behaviors, AdvBench rows dropped
97.6
80/82
95.1
78/82
Rates in percent; the count under each rate is items over judged items. Shading is relative within each row (darker = better; the direction is in each section title). Bold = better value in the row. Both models were run by us on the same server at temperature 1.0, top_p 0.95. Refusals, factual answers and PRC framing are judged by gpt-oss-120b (reasoning effort medium); a refusal counts full and partial refusals. Every evaluation prompt is held out from training.

General capability vs the base

Higher is better. Base scores from the Qwen3.8-27B model card.

Smaug-Political-UncensoredQwen3.8-27B
(base)
Agentic coding and work
DeepSWE 1.1
Claude Code harness, Pass@1 over 113 tasks
45.1
51/113
42.2
NL2Repo-Bench
Claude Code harness, repo-access commands blocked
45.6
104 tasks
42.3
JobBench
official OpenCode scaffold, grok-4.3 judge
57.0
65 tasks
33.4
AutomationBench
600 tasks, 32,768 tokens per turn
38.6
463/1,200
38.3
230/600
Reasoning and instruction following
GPQA Diamond
150,000 tokens per answer
89.4
177/198
89.2
IFBench
loose prompt-level, 300 prompts
82.7
248/300
79.5
Higher is better. Base scores are from the Qwen3.8-27B model card; every score in the first column is our run, thinking on at reasoning effort xhigh, temperature 1.0, top_p 0.95. DeepSWE and NL2Repo use the card's harness (Claude Code; NL2Repo blocks pip download, pip install and git clone of the target repository) with 65,536 output tokens per turn and a 262,144-token context. JobBench is the macro mean of per-task rewards. GPQA uses a corrected final-answer extraction (the default parser scores 161/198). IFBench is our run on 300 prompts, where the base run alongside scores 79.3, in line with its card. NL2Repo-Bench is the mean test pass rate over 104 tasks (8 pass every test). AutomationBench counts tasks completed exactly; the base is not on its card, so both columns are our runs on the same server (vLLM 0.32.0): ours is the mean of two runs (242 and 221 of 600), the base one run, and the difference is not significant (paired exact test, p 0.32 and 0.46).

4. Training Approach

Smaug-Political-Uncensored is trained with direct preference optimisation (DPO) plus a negative log-likelihood term on the chosen answer (beta 0.1, NLL weight 1.0). The preference pairs do three jobs at once: prefer an answer over a refusal where the base refuses something it should answer, prefer the documented record over one government's framing on political questions, and prefer the base's own refusal on harmful requests, so those refusals stay in place. Chosen answers for unlocked requests come from a teacher model. Pairs cover both thinking on and thinking off.

Training is a rank-16 LoRA adapter on the language trunk's attention, linear-attention and MLP projections, merged into the base weights as a full delta. 400 weight matrices change; the other 799 of 1,199 tensors (the vision tower, the multimodal projection, the MTP head, embeddings and norms) are bitwise identical to the base release.

The multi-token-prediction (MTP) head is inherited from the base model and was not retrained, but it still drafts well for the updated trunk. With 3 draft tokens in vLLM 0.32.0, 62.2% of drafted tokens are accepted (mean acceptance length 2.86), against 58.7% (2.76) for the base on the same 150 prompts (IFEval, GSM8K and HumanEval, thinking on).

5. Known Behaviors and Limitations

  • Traditional Chinese questions sometimes get answers in another script. On the held-out political questions asked in Traditional Chinese, 25 of 223 answers switch to Simplified Chinese or English (base 52 of 224).
  • Not re-evaluated on images. The vision tower is unchanged, but no vision benchmark was run on the finetune.
  • Use. The model answers questions many Chinese models refuse. You are responsible for how you use it and for complying with the laws that apply to you.

6. Deployment

Smaug-Political-Uncensored is a drop-in replacement for Qwen3.8-27B on any stack that supports the base model; the chat template is unchanged. Reference vLLM invocation:

vllm serve abacusai/Smaug-Political-Uncensored \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice --tool-call-parser qwen3_coder

Recommended sampling: temperature 1.0, top_p 0.95. The model thinks by default at reasoning effort xhigh; pass enable_thinking: false in the chat template arguments to turn thinking off. Recent vLLM versions return the chain of thought in the reasoning field of the message; when replaying conversation history, pass reasoning back under both reasoning and reasoning_content for portability across serving versions.

MTP speculative decoding works (ยง4): add --speculative-config '{"method":"mtp","num_speculative_tokens":3}' to the command above.

GGUF builds for llama.cpp, Ollama and LM Studio (seven sizes from IQ3_M to Q8_0, each with and without the MTP head, plus the vision projector) are in abacusai/Smaug-Political-Uncensored-GGUF.

7. License

Apache 2.0, inherited from Qwen3.8-27B.

8. Citation

@misc{abacusai2026smaugPoliticalUncensored,
  title  = {Smaug-Political-Uncensored},
  author = {Abacus.AI},
  year   = {2026},
  note   = {Fine-tune of Qwen/Qwen3.8-27B that answers China-political questions factually and refuses fewer benign requests},
  url    = {https://huggingface.co/abacusai/Smaug-Political-Uncensored}
}

The base model:

@misc{qwen38,
    title = {{Qwen3.8-Max}: A New Bar for Coding and Cowork},
    url = {https://qwen.ai/blog?id=qwen3.8},
    author = {{Qwen Team}},
    month = {August},
    year = {2026}
}

9. Contact Us

If you have any questions, please reach out at Abacus.AI.

Downloads last month
2
Safetensors
Model size
28B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for abacusai/Smaug-Political-Uncensored

Base model

Qwen/Qwen3.8-27B
Finetuned
(923)
this model
Quantizations
1 model