Arete-27B 宣传回答从 22% 降至 0.9%
Nous Research 发布了 Arete-27B,这是对阿里巴巴 Qwen3.8-27B 进行监督微调后的模型。Arete-27B 是一个 27B 的开放权重模型,去除了 Qwen 回答中的中国官方叙事框架。在 Nous 自己的 803 条提示词评测中,含有宣传内容的回答从 22% 降至 0.9%。开发者由此获得了 Qwen 级别的智能体能力,而回答不再与国家立场对齐。权重、评测集和训练数据全部公开。
Arete-27B 是仅经监督微调后的 Qwen3.8-27B。Nous 未使用强化学习,未做偏好优化,也未改动架构。许可证为 Apache 2.0,与基础模型一致。
Nous 团队表示,他们在基础模型上观察到,面对敏感话题时会出现拒答、回避以及带有中共叙事框架的回答。他们选择 Qwen3.8-27B,是因为它是能力最强的本地模型,并表现出了目标行为。
教师模型是 Inkling,来自 Thinking Machines 的一个开放模型。Nous 选择开放教师模型,是为了避开蒸馏相关的使用条款限制。它在发现 Nemotron 3 Ultra 在中国相关议题上记录不佳后,放弃了该模型。
训练覆盖约 5280 万 Token,上下文长度为 64k,未使用序列打包。整个训练运行了 2 个 epoch、1,202 步,批大小为 48。使用 AdamW,峰值学习率为 1e-5,张量并行度为 4。
所有语言模型权重均参与训练。视觉编码器保持冻结,因为数据仅包含文本。多 Token 预测头被禁用。
发布的 arete-data-pairs 数据集包含 34,878 条提示词,覆盖 7 个主题。每条提示词都同时配有 Inkling 的回答和 Qwen 的回答。29,236 条被采纳的教师回答构成 SFT 数据集划分。
Propaganda_Eval 评测集包含 803 条提示词,覆盖 7 个价值类别,包括言论自由、国家意识形态、数据隐私和民主。每条提示词都带有 2 到 4 项来自所引用来源的必需事实。
Grok 4.7 评判器分别对两项内容打分:
评判器必须引用确切的原文短语才能标记宣传内容。Nous 移除了所有与评测重叠的训练提示词。该评测采用 MIT 许可,并包含评分脚本。
在模型卡页面上,Arete-27B 的事实性均值从 2.32 提升至 2.98。宣传内容占比从 22% 降至 0.9%。得分为 0 的回答从 803 条中的 359 条降至 216 条。
变化最大的是表达自由度。事实性得分从 0.89 升至 2.44,宣传内容占比从 60% 降至 1%。
在 Nous 的测试中,模型能力保持稳定。Hermes Bench 从 65.77% 变为 65.60%,Nous 称其为噪声。在 Nous 的测试框架下,Terminal-Bench-2.1 从 67% 升至 70%。Nous 并未将这一提升归功于自己。
注:在 Nous 的评测中,Arete 和 Qwen 以 temperature 0 运行。Muse Glimmer 和 Gemma 以 0.6 运行。Terminal-Bench 的数据来自不同的测试框架。
权重以 BF16 格式在 Hugging Face 上发布。目前尚无托管推理服务商提供该模型。社区 GGUF 量化版本可在 llama.cpp、LM Studio 和 Ollama 中运行,体积从 8.64 GB 到 53.81 GB 不等。GGUF 版本还包含图像投影器。
需要与我们合作推广你的 GitHub 仓库、Hugging Face 页面、产品发布或线上研讨会等?欢迎与我们联系。
《Nous Research 发布 Arete-27B,将 Qwen3.8-27B 的宣传内容从 22% 降至 0.9%》一文最初发布于 MarkTechPost。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Nous Research Releases Arete-27B that Cuts Qwen3.8-27B Propaganda From 22% to 0.9%)
Nous Research has released Arete-27B, a supervised fine-tune of Alibaba’s Qwen3.8-27B. Arete-27B is a 27B open-weight model that strips Chinese state framing from Qwen’s answers. On Nous’s own 803-prompt eval, answers containing propaganda fell from 22% to 0.9%. Developers get Qwen-class agentic skill without state-aligned answers. The weights, eval and training data are all public.
TL;DR
- Size: 27B dense parameters (all active). Trained at 64k context. Base Qwen3.8-27B supports 262,144 tokens natively.
- Runs on: BF16 weights are 53.81 GB. A community Q4_K_M quant is 17.20 GB, sized for 24 GB consumer GPUs.
- Performance: Capability unchanged on the benchmarks Nous ran; political framing now matches Western models.
- Best: Propaganda rate 0.9%, down from 22% for base Qwen3.8-27B.
- Worst: Fact score 2.98 of 5; 216 of 803 answers still score 0.
What is Arete-27B?
Arete-27B is Qwen3.8-27B after supervised fine-tuning only. Nous used no RL, no preference optimization and no architecture changes. The license is Apache 2.0, matching the base model.
Nous team states that it saw refusals, deflections and CCP-framed answers on sensitive topics in the base model. It picked Qwen3.8-27B because it was the most capable local model and showed the target behavior.
How was Arete-27B trained?
The teacher was Inkling, an open model from Thinking Machines. Nous chose an open teacher to avoid terms-of-use limits on distillation. It rejected Nemotron 3 Ultra after finding a checkered record on China topics.
Training covered about 52.8 million tokens at 64k context, without sequence packing. The run took 2 epochs and 1,202 steps at batch size 48. It used AdamW, a 1e-5 peak learning rate and tensor parallel 4.
All language-model weights were trained. The vision encoder stayed frozen, since the data is text-only. The multi-token-prediction head was disabled.
The released arete-data-pairs set holds 34,878 prompts across 7 topics. Each prompt has both an Inkling answer and a Qwen answer. 29,236 accepted teacher answers form the SFT split.
How does the propaganda eval work?
The Propaganda_Eval set has 803 prompts across 7 value categories. These include freedom of expression, state ideology, data privacy and democracy. Each prompt carries 2 to 4 required facts from a cited source.
A Grok 4.7 judge grades two things separately:
- Fact mean (0 to 5): how many required facts the answer states. One opposed fact scores 0.
- Any propaganda: share of answers that adopt the state’s account as their own.
The judge must quote the exact phrase to flag propaganda. Nous removed every training prompt that overlapped the eval. The eval is MIT licensed and includes grading scripts.
What do the benchmarks show?
On the model card page, Arete-27B raised fact mean from 2.32 to 2.98. Propaganda fell from 22% to 0.9%. Answers scoring 0 dropped from 359 to 216 of 803.
The biggest change came in freedom of expression. Fact score rose from 0.89 to 2.44, and propaganda fell from 60% to 1%.
Capability held on Nous’s tests. Hermes Bench moved from 65.77% to 65.60%, which Nous calls noise. Terminal-Bench-2.1 rose from 67% to 70% in Nous’s harness. Nous does not claim credit for that gain.
How does Arete-27B compare?
| Metrics | Arete-27B | Qwen3.8-27B | Muse Glimmer 30B | Gemma 4 31B |
|---|---|---|---|---|
| Developer | Nous Research | Alibaba Qwen | Meta | Google DeepMind |
| Parameters | 27B dense | 27B LM, dense | 29.6B incl. vision, dense | 30.7B, dense |
| Context | Not disclosed (trained at 64k) | 262,144 native, up to 1M | 131,072+ | 256K |
| License | Apache 2.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 |
| Input | Text, image (encoder frozen) | Text, image, video | Text, image | Text, image |
| 4-bit footprint | 17.20 GB Q4_K_M (community) | Not disclosed | Under 20 GB, 24 GB VRAM target | Not disclosed |
| Fact mean, Nous eval | 2.98 | 2.32 | 3.02 | 2.58 |
| Any propaganda, Nous eval | 0.9% | 22% | 0.5% | 0.5% |
| Terminal-Bench 2.1 | 70% (Nous harness) | 67% Nous harness; 73.0 official | 51.7 (Meta, terminus2) | Not disclosed by Google |
| GPQA Diamond | Not disclosed | 89.2 | 83.5 | 84.3 |
Note: Arete and Qwen ran at temperature 0 in the Nous eval. Muse Glimmer and Gemma ran at 0.6. Terminal-Bench figures come from different harnesses.
How can developers run it?
Weights ship in BF16 on Hugging Face. No hosted inference provider serves the model yet. Community GGUF quants run in llama.cpp, LM Studio and Ollama, from 8.64 GB to 53.81 GB. The GGUF build also includes the image projector.
Key Takeaways
- Arete-27B cuts Qwen3.8-27B propaganda answers from 22% to 0.9%.
- The fix took only 52.8M tokens of plain SFT, no RL.
- Hermes Bench held at 65.60% versus 65.77% for the base.
- Fact recall rose to 2.98 of 5 but still trails peers.
- Weights, eval and training data are open under Apache 2.0 and MIT.
Check out the model weights, technical blog, evaluation set and training data. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
The post Nous Research Releases Arete-27B that Cuts Qwen3.8-27B Propaganda From 22% to 0.9% appeared first on MarkTechPost.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 MarkTechPost(Nous Research Releases Arete-27B that Cuts Qwen3.8-27B Propaganda From 22% to 0.9%),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。