Arete-27B: Based on Qwen3.8-27B, propaganda responses reduced from 22% to 0.9%

A model post-trained on Alibaba's Qwen3.8-27B to reinforce Western values. It focuses on correcting political biases inherent in models trained in China, such as refusal responses, evasion, or answers framed in specific ways.

Evaluation results showed the factuality score increased from 2.32 to 2.98 on a 5-point scale, and the proportion of responses containing propaganda components dropped sharply from 22% to 0.9%. These improvements were consistent across all 7 value evaluation items.

Released under the Apache 2.0 license, it is provided as a 27B parameter model in BF16 format safetensors files. It is suitable for developers looking to address the value fixation issues present in existing open-source models.