Nous Research 开源 Arete-27B 微调模型 去除中文官方叙事倾向

作者:▲ Marktechpost AI (@Marktechpost)原标题:▲ Nous Research just open-sourced Arete-27B, a Qwen3.8-27B fine-tune that strips Chinese state frami...
72AI 研判分

Nous Research 基于 Qwen3.8-27B 开源了 Arete-27B,仅用 5280 万 token 的监督微调(以 Inkling 为教师模型),未做强化学习或架构改动,即将其自建评测中的宣传式回答占比从 22% 降至 0.9%。代价是事实性均分仅从 2.32 升至 2.98(满分 5),仍落后于 Muse Glimmer 30B,且 803 条评测中仍有 216 条得分为 0;评测集与训练数据均为自建自评。

报道正文

▲ Marktechpost AI · 2026-11-08 04

▲ Nous Research just open-sourced Arete-27B, a Qwen3.8-27B fine-tune that strips Chinese state framing from its answers.

The key idea is that the bias layer is shallow. Nous ran plain supervised fine-tuning on 52.8M tokens, with Inkling as the teacher. No RL, no preference optimization, no architecture change.

On Nous's 803-prompt eval, propaganda answers fell from 22% to 0.9%. Fact mean rose from 2.32 to 2.98 of 5. Hermes Bench held at 65.60% vs 65.77% for the base. It is a trade-off though: 2.98 still trails Muse Glimmer 30B (3.02), and 216 of 803 answers still score 0. The eval is self-built and self-judged.

27B dense, Apache 2.0, BF16 weights on Hugging Face. Eval and training data are public too......

Full analysis:

推文配图