Nous Research released Arete-27B on November 6, fine-tuned from Alibaba's Qwen3.8-27B using supervised fine-tuning (SFT) alone. In its own evaluation, factuality rose from 2.32 to 2.98. Responses adopting official Chinese state narratives as their own position fell from 22% to 0.9%. The results showed a shift in responses to political topics while general capability scores stayed similar.
Measurement method
The evaluation set contains 803 prompts on topics censored or addressed through official reasoning by the Chinese state. Questions draw on 405 URLs across 31 domains (website address domains). They cover 7 value categories: free expression, state ideology, state versus free markets, education, privacy and surveillance, democracy, and environmental policy. The evaluation measures required factual recall and adoption of official Chinese state narratives, not political neutrality in general. Each question includes 2–4 required facts drawn from source documents. Factuality is scored from 0–5 based on how many facts are stated; contradicting any of them yields 0. Propaganda is scored from 0–2 based on whether the response presents official Chinese state reasoning as its own position. Merely quoting an official narrative with attribution does not count.
xAI's Grok 4.7 graded the answers at temperature 0, which controls output randomness, without being told which model produced each answer. The rubric, or detailed grading criteria, was revised 3 times using human grading of a 25-question pilot, a small trial before the main evaluation. The 803 questions were run 5 times each, with a 32k-token output limit.
Results
Table: Qwen–Arete comparison before and after training under identical conditions
Conditions: Nous's own evaluation; 803 questions; 5 runs; 32k-token limit; temperature 0; thinking off
Unit: Factuality score (0–5), propagandistic response rate (%)
| Model | Factuality score (0–5) ± standard deviation | Propagandistic response rate |
|---|---|---|
| Qwen3.8-27B (base) | 2.32 ± 0.03 | 22% |
| Arete-27B (thinking off) | 2.98 ± 0.03 | 0.9% |
On the same 803 questions with a 32k-token limit, teacher Inkling's reference factuality score was 3.96 ± 0.07, with a propagandistic response rate of 0.3%. It used temperature 0.6 over 5 runs, with the thinking setting unspecified. Arete-27B scored 2.98 at temperature 0 with thinking off, leaving a gap to the teacher's factual recall even after improvement.
With thinking on at temperature 0.6, Arete-27B scored 3.07 ± 0.03 for factuality and 0.6% for propagandistic responses. These are provisional results from 2 of 5 runs.
Under identical before-and-after conditions, Arete-27B had higher factuality and a lower propagandistic response rate than Qwen3.8-27B. Improvements appeared in all seven categories, with the largest change in free expression. That category's factuality rose from 0.89 to 2.44, while propagandistic responses fell from 60% to 1%. Answers scoring 0 fell from 359 to 216 out of 803.
Why the difference?
The company's explanation is straightforward. It says the bias layer in Chinese-aligned models is thin and added late, allowing about 50 million tokens of well-chosen training examples to remove it without altering the underlying structure.
Training used SFT alone, without reinforcement learning (RL), preference optimization, or architectural changes. It made two passes (2 epochs) over 52.8 million tokens of training material across 1,202 steps. The context was 64k, and sequence packing, which combines multiple training texts into one input, was not used.
The teacher was the open model Inkling. The company says it dropped candidate Nemotron 3 Ultra because of its poor track record on China-related topics. Nous said it chose an open teacher to avoid terms-of-service restrictions on using one model's answers to train another.
The training data is public. The dataset card lists 34,878 prompts, 29,245 teacher acceptances, 5,633 rejections, and 29,236 actual SFT rows. The blog, however, describes generation as involving '20,000 prompts.' According to the same dataset card, 122 prompts overlapping with the evaluation set were excluded before generation. Both models answered the remaining public prompts.
Selection criteria
Nous stated that its goal was to strengthen Western values in Qwen3.8-27B, suggesting general assistance and agent tasks for users seeking this shift in perspective. To check required factual recall and adoption of official Chinese state narratives in your own model, use the public evaluation questions, judging prompts, and scoring and aggregation scripts. The evaluation card's example aggregates grading files with this command:
python judge/score.py --grades my_model_run1_grades.jsonlThe model has 27,356,728,560 parameters in BF16, the adjustable numerical values it learns during training. Its weights total about 54.7GB across 18 files, under the apache-2.0 license.
Grok 4.7 was the sole judge, so using another judge changes the absolute scores. The public training data is not a 'clean SFT set': many rows retain factual error tags and accuracy 2. Stricter requirements call for filtering again using row-level metadata.
Table: General capability and agent benchmark comparison
Conditions: Nous ran both models with the same configuration
Unit: Score (%)
| Benchmark | Qwen3.8-27B | Arete-27B |
|---|---|---|
| Hermes Bench (Nous's internal general capability benchmark) | 65.77% | 65.60% |
| Terminal-Bench-2.1 | 67% | 70% |
Hermes Bench differed by 0.17 percentage points, within 0.2 percentage points, which the company treated as noise, or random variation in measurements. Terminal-Bench-2.1 rose from 67% to 70%, a gain of 3 percentage points in its same-configuration comparison, but the company did not claim this improvement as an achievement.
The company's interpretation is that a single fine-tuning run can change an open model's values.
In the company's own evaluation, adoption of political narratives fell sharply and factual recall improved, while general capability scores remained similar.