Skip to content- Sources: primary, discussion
- Summary: Nous Research published Arete-27B under Apache 2.0, trained by supervised fine-tuning only on about 52.8 million tokens of answers from the open Inkling model over two epochs and 1,202 steps at 64k context. On an 803-prompt evaluation graded by a Grok 4.7 judge, the fact score rose from 2.32 to 2.98 out of 5 and answers containing state framing fell from 22% to 0.9%. Hermes Bench moved from 65.77% to 65.60% and Terminal-Bench-2.1 from 67% to 70%, which Nous does not credit to the training.
- Why it matters: It is a fully specified recipe, with optimizer settings, data size and a held-out eval, for shifting an open-weight model's values with SFT alone and no measured capability loss.
- Background: Supervised fine-tuning trains a model on prompt and response pairs, here written by a stronger teacher model, which is a form of distillation. Nous chose an open teacher to avoid the terms-of-use limits that closed providers place on training from their outputs. The model card separates factual recall from framing, so a model can score well on facts while still adopting an official line.