Nous Research Fine-Tunes Qwen Model to Cut State Propaganda Rate Below 1%

1 reports archived1 independent sourcesupdated 11/7/2026, 20:32:45

Synthesis & Latest Updates

1 Sources Cross-Validated

Nous Research supervised fine-tuned Alibaba's Qwen3.8-27B on 53 million tokens to produce Arete-27B, cutting the rate of state-propaganda responses from 22% to under 1% while preserving general capabilities. The team argues this shows political bias in Chinese open-weight models is only a superficial layer, and has released the weights, test set, and training data under Apache 2.0.

LATEST/Nous Research supervised fine-tuned Alibaba's Qwen3.8-27B on 53 million tokens to produce Arete-27B, cutting the rate of state-propaganda responses from 22% to under 1% while preserving general capabilities. The team argues this shows political bias in Chinese open-weight models is only a superficial layer, and has released the weights, test set, and training data under Apache 2.0.

Hourly heat curve

2 fully observed hours

Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first

The DecoderT2Nov 7, 20:32·71 pts
Nous Research Fine-Tunes Qwen Model to Cut State Propaganda Rate Below 1%
- Supervised fine-tuning on 53M tokens cut propaganda rate from 22% to under 1% with little loss in general capability
- Authors argue political bias in Chinese open-weight models is surface-level alignment rather than deeply embedded
- Weights, test set, and training data released under Apache 2.0 for third-party replication and evaluation
View Related Reports & Analysis Original Source