Oriel — AI & technology worth noticing, every day.

NousResearch/Arete-27B

AI model · https://huggingface.co/NousResearch/Arete-27B

First seen by Oriel on .

AI summary

NousResearch describes Arete-27B, an SFT post-training of Alibaba's Qwen3.8-27B that reduced propaganda-framed answers on its evaluation while retaining similar general and agentic benchmark scores.

  • Training used about 52.8 million text tokens, 20,000 prompts, a 64k context, two epochs, and 1,202 steps, with Inkling as the teacher.
  • Across 803 prompts, fact mean rose from 2.32 to 2.98 out of 5, while answers containing propaganda fell from 22% to 0.9%.
  • The improvement appeared in all seven value categories; freedom of expression had the largest reported change, with fact score rising from 0.89 to 2.44 and propaganda falling from 60% to 1%.
  • Hermes Bench was nearly unchanged at 65.60% versus 65.77%, while Terminal-Bench-2.1 scored 70% versus 67%; the authors do not claim the latter as an improvement.
  • The authors note that the evaluation uses an LLM judge and that decoding settings were not identical across all compared models.

#llm fine-tuning #value alignment #model evaluation #qwen

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 675, latest observation ; +570 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 9,187, latest observation ; +9,187 over observed period since

Aliases

SourceExternal IDURL
Hugging Face ModelsNousResearch/Arete-27Bhttps://huggingface.co/NousResearch/Arete-27B Read discussion