Nous Research Open-Sources Arete-27B Fine-Tune with Evaluation Data
Built via supervised fine-tuning on Qwen3.8-27B and released under Apache-2.0, the model is accompanied by an open evaluation set and methodology for measuring model behaviour.
Nous Research has released Arete-27B, an open-weight fine-tune of Alibaba's Qwen3.8-27B produced entirely through supervised fine-tuning, without reinforcement learning or architectural changes. The model weights, an 803-prompt evaluation set, and the data-composition assets are published under the Apache-2.0 license on Hugging Face.
According to measurements reported by Nous Research, using an LLM-based judge on its own evaluation set, answers exhibiting propaganda-style patterns fell from 22% to 0.9%, while general-capability benchmark results remained broadly unchanged. The company states the fine-tune used roughly 53 million tokens over two epochs, with training prompts generated by the open-source Inkling model. These figures are vendor-reported and have not been independently verified.
The release suggests that behavioural alignment in open-weight models can be comparatively shallow and adjustable through modest supervised training. Beyond the model itself, the published evaluation set and methodology provide an open approach to measuring model behaviour, which may be relevant for organisations assessing open-weight models for regulated or customer-facing deployments.