最新模型
Nous Releases Arete-27B, Using Supervised Fine-Tuning to Adjust Qwen’s Answers on Political Issues
Nous Research has released Arete-27B’s weights, training data, and evaluation set, demonstrating a way to use supervised fine-tuning to change how Qwen3.8-27B responds. The team reports improvements on specific evaluations and broadly unchanged agent capabilities, but the results remain subject to limits in prompt design, judge models, and decoding settings.
Nous Research announced Arete-27B on November 6. Starting from Qwen3.8-27B, the model uses supervised fine-tuning to adjust its responses to politically sensitive issues. The release includes open weights, training data, and an evaluation set, allowing researchers to inspect how the behavior changed and rerun the tests themselves. (Release announcement, official description)
Training used about 52.8 million tokens with a 64k context length, over two epochs and 1,202 steps; no reinforcement learning or preference optimization was added. The team used Inkling as the teacher model, updated all language-model weights, froze the vision encoder, and disabled the multi-token prediction head. These settings define the scope of the experiment: it primarily tests how text supervision can change response behavior, while multimodal capabilities still need to be validated separately. (Model card)
The evaluation covered seven issue categories and 803 prompts. Each prompt listed its source and two to four required facts, and Grok 4.7 graded the responses without knowing which model produced them. Nous reported that, in a comparison with thinking disabled and temperature set to zero, the factuality score rose from 2.32 to 2.98 on a five-point scale. The share of answers classified by its rules as adopting the Chinese government’s official narrative fell from 22% to 0.9%. These results use metrics defined by the team and should not be read directly as measures of general factual accuracy or political neutrality. (Evaluation data)
On capability retention, the official internal Hermes Bench score was 65.60%, compared with 65.77% for the base model; Terminal-Bench-2.1 scores were 70% and 67%, respectively. The team cautioned that its inference and testing framework settings may differ from those used on official leaderboards, and did not present the latter result as a definitive improvement. For engineering teams, this is a reproducible example of behavioral adjustment, but it is not yet enough to show that all capabilities were retained. (Official results)
Next, researchers should look for independent replications and results on Chinese prompts, with different judges, and on multimodal tasks. The evaluation data is labeled as English, and comparisons across models did not all use the same decoding settings; Chinese-language products therefore still need their own test sets. The research value lies in making inspectable training and scoring materials public. The explanation that the “bias layer is shallow” remains Nous’s interpretation of this experiment. (Model card and limitations)