Nous Research Releases Arete-27B: Propaganda Rate Drops from 22% to 0.9%

The open-source model fine-tuning space has produced another noteworthy case. Nous Research recently released Arete-27B, an Apache 2.0-licensed fine-tuned model built on Qwen3.8-27B. Its most striking figure comes from Nous's own 803-prompt evaluation: the proportion of propaganda-style responses fell from 22% in the base model to 0.9%, while the Hermes Bench score held steady at 65.60%, essentially unaffected.

A "subtractive" fine-tune: cutting propaganda while preserving general capability

Fine-tuning large models usually aims at capability enhancement—stronger reasoning, longer context, more accurate instruction following. Arete-27B takes a different path: targeted weakening of a specific class of output.

According to information published by Nous Research, on the 803-prompt evaluation set, 22% of responses from the base model Qwen3.8-27B were judged to be propaganda in nature; after Arete-27B fine-tuning, this proportion dropped to 0.9%. That said, the Hermes Bench score remained unchanged from the base's 65.60%, indicating that this fine-tune did not come at the cost of general conversational and instruction-following capability.

This "subtract-only, no-add" fine-tuning approach is uncommon in the open-source community. Most publicly released fine-tuned models report improvements across multiple benchmarks at once, whereas Arete-27B's core selling point is a declining metric. This precisely reflects Nous Research's understanding of "alignment": not making the model more talkative, but making it converge where it should.

---

Parameter-efficient fine-tuning (PEFT) is currently the mainstream approach for secondary training of open-source models, with LoRA (Low-Rank Adaptation) being the most representative. Its method is to freeze all original weights of the base model and attach a pair of low-rank matrices in parallel beside certain linear layers, updating only this small pair during training. Since the trainable parameter count typically accounts for only one-thousandth to one-hundredth of the full model, LoRA's requirements for VRAM and compute are far lower than full-parameter fine-tuning, and the artifact is merely an adapter weight of tens to hundreds of MB, easy to distribute and stack. The trade-off is limited capacity, which may prove inadequate for tasks requiring substantial changes in model behavior. In contrast to LoRA is full-parameter fine-tuning (FFT), which updates all weights—expressive but costly, and prone to forgetting base capabilities. Arete-27B did not disclose which route it took, which is one of the key gaps making its method hard to reproduce.

Why the propaganda metric deserves separate attention

"Propaganda" in model evaluation typically refers to output that carries strong bias, is inflammatory, or serves a particular stance. Such output does not necessarily violate safety policies, yet it can affect the model's neutrality as an information tool.

For scenarios requiring objective analysis, factual synthesis, or neutral summarization, a 22% propaganda response rate means nearly a quarter of responses may carry implicit stance drift. Pressing it down to 0.9% has direct value for downstream applications: reducing post-processing filtering costs and improving output credibility.

It should be noted that both the 22% and 0.9% figures come from Nous's own 803-prompt evaluation, not an independent third-party benchmark. The specific composition of the evaluation set, the criteria for judging propaganda, and inter-annotator agreement were not elaborated in the original text. This means the data is better treated as directional reference rather than an absolute performance commitment.

---

Inter-annotator agreement is a core metric for measuring the reliability of subjective evaluations, commonly quantified using Cohen's kappa or Krippendorff's alpha. When judgment criteria involve subjective assessments such as "whether it carries bias," different annotators may classify the same response quite differently. If agreement is low, the gap between 22% and 0.9% could partly stem from annotation noise rather than a genuine change in model behavior. Typically, kappa above 0.6 counts as substantial agreement, and above 0.8 as near-perfect agreement. The original text did not disclose the number of annotators, the agreement coefficient, or the detailed judgment rules for this evaluation, so readers should apply appropriate discounts when citing these figures.

Apache 2.0 license and the choice of Qwen base

Arete-27B uses the Apache 2.0 license with Qwen3.8-27B as its base. Both choices are worth noting.

Apache 2.0 permits commercial use, modification, and redistribution, making it relatively friendly to teams hoping to integrate the model into product pipelines. Choosing Qwen3.8-27B as the base continues the open-source community's trend of using mid-scale models for targeted fine-tuning—a 27B parameter count offers a compromise between inference cost and capability, suitable for single-machine multi-GPU or mid-scale deployment.

Nous Research is well known in the open-source model community for its Hermes series of fine-tunes, and Hermes Bench is also its own evaluation system. Arete-27B's Hermes Bench score matches the base, indicating that the fine-tuning data and training strategy did not disrupt the base's original instruction-following distribution.

---

The key difference between Apache 2.0 and GPL-style copyleft licenses is that it does not require derivative works to be open-sourced under the same license. Users can modify the model and integrate it closed-source into commercial products, provided they retain copyright and license notices and indicate modifications. However, it does not grant trademark rights and offers no additional guarantees beyond patent retaliation clauses. By comparison, the community license used by the Llama series carries additional restrictions such as monthly active user thresholds, while some models restrict commercial use. For teams wanting to embed a model into product pipelines without being bound by open-source obligations, Apache 2.0 is one of the least restrictive options—which is why Arete-27B is relatively attractive on the licensing front.

Fine-tuning method undisclosed: reproducibility in question

Currently, public information stops at the results level. Nous Research did not disclose the training data composition, fine-tuning method (whether full-parameter fine-tuning or a parameter-efficient method like LoRA), training steps, or the specific judgment process for the propaganda evaluation.

For developers wanting to reproduce or further fine-tune on this basis, the absence of this information introduces uncertainty. The Apache 2.0 license grants freedom of use, but "how it was done" remains a black box. The community's next focus will likely be on whether a third party independently verifies the 803-prompt evaluation set, and whether the drop in propaganda rate holds across other prompt distributions.

Targeted alignment may become a new direction for open-source fine-tuning

The Arete-27B case suggests a possibility: open-source fine-tuning need not always pursue across-the-board benchmark gains—precise reduction along a single behavioral dimension can likewise constitute a valuable product.

The drop from 22% to 0.9% is striking enough in numbers, but the real test lies in generalization—whether, outside Nous's own evaluation set, the model similarly reduces biased output without becoming overly conservative or prone to refusals. The unchanged Hermes Bench is a positive signal, but a single benchmark cannot cover the entire behavioral space.

For developers, Arete-27B offers a commercially usable 27B option with a markedly lower propaganda tendency; for researchers, it raises a question worth tracking: where are the boundaries of targeted behavioral fine-tuning, and how can more transparent evaluation standards be established to verify such "subtractive" effects?

One-line Summary

Nous Research releases Arete-27B, using "subtractive" fine-tuning to cut propaganda output rate from 22% to 0.9%.

Paragraph Summary

Nous Research has released Arete-27B, an Apache 2.0-licensed fine-tuned model based on Qwen3.8-27B, focused on targeted weakening of propaganda-style output. In its own 803-prompt evaluation, the proportion of such responses fell from the base's 22% to 0.9%, while the Hermes Bench score held at 65.60%, indicating general instruction-following capability was not sacrificed. The article argues that this "subtract-only" fine-tuning approach is rare in the open-source community, and its value lies in reducing downstream post-processing costs and improving output neutrality. However, the evaluation set composition, propaganda judgment criteria, training data, and method are all undisclosed, leaving reproducibility and generalization to be verified by third parties.