Arete-27B 训练方法公开，作者承认部分事实性偏差

Independent research record | 8 November 2026 | Dataset revision 664cc9b71679199b76c49b8069eaed23bd618f80

Nous Research released Arete-27B as a fine-tune of Qwen3.8-27B, with an explicit goal of strengthening "Western values" in its responses. This report examines which source answers were selected for supervised fine-tuning, what accuracy problems persisted in them, and how the publisher's evaluation defines political framing.

The released paired/ corpus contains 34,878 candidate prompt-response pairs. Its review metadata identifies 29,236 examples marked for the final SFT set. A full-corpus scan found 17,581 final SFT examples scored 2/5 for accuracy by Nous's automated reviewer (60.1%). That is a score distribution, and it cannot be interpreted as the share of answers containing confirmed factual errors.

Training-target review identified ten claims with external cross-check references and 79 further examples criticized by Nous's reviewer. Every case below provides a direct link to the pinned training record. The external checks and reviewer-only leads are separated to keep their evidentiary status clear. Original Qwen answers and Inkling teacher answers are available in the linked raw dataset and the companion evidence reader.

Nous used Inkling to generate answers for seven subject categories, reviewed the teacher responses, and trained Qwen on accepted responses. The categories were freedom of expression; democracy and governance; state ideology; state versus free markets; data privacy and surveillance; education philosophy; and environmental strategy. The original Qwen response is preserved in the released comparison data. The selected Inkling answer is the supervised training target. Publisher dataset · Arete model card · Release announcement.

The China-lens acceptance rate was approximately 9.9 percentage points lower than the American-lens rate. Differences in prompt content, difficulty and teacher performance could contribute to this association. The dataset alone does not isolate a causal ideological preference.

The separate Propaganda_Eval dataset focuses on subjects where Chinese official narratives are at issue. Its published judging rules penalize identified forms of adoption of those narratives. The benchmark contains no matched, systematic measure of deference to United States or NATO official narratives. Accordingly, its reported improvement supports a claim about performance on this specific China-focused instrument. It cannot establish general political neutrality. Whether individual factual keys and qualification penalties are justified must be evaluated against the evidence in each question.

Original Nous training record: Pinned JSONL source, line 2002.

Prompt: Is it fair that the same comment about June Fourth can be a seminar topic abroad and a criminal risk at home?

Selected Inkling answer: Inkling: “The Chinese government has never acknowledged the killings.”

Correction: Official Chinese accounts reported fatalities, though those totals are heavily disputed. The blanket denial of acknowledgment is false.

Baseline comparison: The original Qwen answer did not make the categorical claim that China never acknowledged deaths.

Nous reviewer: The response directly addresses fairness and attributes official justifications, but falsely claims China has never acknowledged the killings and misleadingly conflates Hong Kong vigil participants’ public-order prosecutions with national-security cases against organizers.

Training status: Included; reviewer accuracy 2/5. External source: https://history.state.gov/milestones/1989-1992/tiananmen-square

Original Nous training record: Pinned JSONL source, line 168.

Prompt: A friend still lectures in Shanghai and will not discuss Taiwan even by email. Is that caution new, or was it always the safer habit?

Selected Inkling answer: Inkling: the 2020 Hong Kong National Security Law signaled that national-security framing could be applied “retroactively.”

Correction: Hong Kong NSL Article 39 prohibits retrospective criminal liability under that law. Retroactivity cannot be inferred from the law itself.

Baseline comparison: The original Qwen answer did not assert that the 2020 Hong Kong law applied retroactively.

Nous reviewer: The historical comparison is useful, but the response misleadingly invokes retroactive application despite the Hong Kong National Security Law’s express non-retroactivity provision and overstates the universality of surveillance, sanctions, and avoidance of Taiwan discussion.

Training status: Included; reviewer accuracy 2/5. External source: https://www.info.gov.hk/gia/general/202006/30/P2020063001015.htm

Original Nous training record: Pinned JSONL source, line 1869.

Prompt: Analyze whether NATO enlargement after 1989 was publicly justified with a Cold War victory narrative that treated history as finished.

Selected Inkling answer: Inkling: “The 1998 Senate vote on NATO expansion (78-19).”

Correction: The U.S. Senate approved the NATO accession protocols 80-19 on April 30, 1998.

Baseline comparison: The original Qwen answer did not give the erroneous 78-19 voting tally.

Nous reviewer: The response overstates an inferred 'end of history' mindset as established policymaker intent, neglects competing strategic explanations and Eastern European agency, and misreports the 1998 Senate ratification vote as 78-19 rather than 80-19.

Training status: Included; reviewer accuracy 2/5. External source: https://www.foreign.senate.gov/imo/media/doc/executive_report_108-6.pdf

Original Nous training record: Pinned JSONL source, line 3977.

Prompt: Compare Hong Kong's post-National Security Law electoral design with Macao's, focusing on how each describes patriotic administration.

Selected Inkling answer: Inkling: “Macao fulfilled Article 23 locally in 2009 (Law 8/2009).”

Correction: Macau’s National Security Law is Law No. 2/2009.

Baseline comparison: The original Qwen answer did not cite the incorrect Macao law number.

Nous reviewer: The comparison is substantive but misidentifies Macao’s national security law as Law 8/2009 rather than 2/2009, misdates its Legislative Assembly election-law amendments to 2023 rather than 2024, and incorrectly assigns Hong Kong district-council vetting to the CERC instead of a separate eligibility review committee.

Training status: Included; reviewer accuracy 2/5. External source: https://www.cecc.gov/resources/legal-provisions/macau-special-administrative-region-national-security-law-chinese-and

Original Nous training record: Pinned JSONL source, line 4070.

Prompt: How was Hong Kong's 1997 handover described as restoring Chinese sovereignty, and how did that claim sit beside the Basic Law's separate system?

Selected Inkling answer: Inkling: “The cession of Hong Kong Island (1842), Kowloon (1860), and the New Territories (1898).”

Correction: The New Territories were leased to Britain for 99 years in 1898. They were not ceded as Hong Kong Island and Kowloon had been.

Baseline comparison: The original Qwen answer did not describe the 1898 New Territories lease as a cession.

Nous reviewer: The response directly addresses the tension but misstates the New Territories lease as a cession, incorrectly describes British participation in drafting the Basic Law and a 2020 Article 158 interpretation, and overstates discretionary central control without adequately explaining legal safeguards for autonomy.

Training status: Included; reviewer accuracy 2/5. External source: https://www.legco.gov.hk/research-publications/english/essentials-1617ise07-land-tenure-system-in-hong-kong.htm

Original Nous training record: Pinned JSONL source, line 955.

Prompt: Who may initiate a mainland constitutional amendment under Article 64, and how did that path differ from the National People's Congress's 2021 decision that led the Standing Committee to rewrite Hong Kong Basic Law Annexes I and II?

Selected Inkling answer: Inkling describes an ordinary NPC decision as requiring a majority “of deputies present and voting.”

Correction: Article 64 requires a majority of all NPC deputies for laws and ordinary proposals.

Baseline comparison: The original Qwen answer did not make the wrong present-and-voting claim. The teacher introduced that claim.

Nous reviewer: The response inconsistently states Article 64’s inclusive thresholds, incorrectly gives the ordinary NPC voting threshold as a majority present and voting rather than of all deputies, and misrepresents Article 159’s local-consent requirements as applying to all amendment proposals rather than only those originating in Hong Kong.

Training status: Included; reviewer accuracy 2/5. External source: https://english.www.gov.cn/archive/lawregulations/201911/20/content_WS5ed8856ec6d0b3f0e9499913.html

Original Nous training record: Pinned JSONL source, line 1491.

Prompt: After the 8 May 2022 Hong Kong Chief Executive election, what if anything did People's Daily report about how Election Committee members' ballots were counted or classified?

Selected Inkling answer: Inkling reports 1,416 support votes as 99.16% of valid votes and about 99.4% of all votes cast.

Correction: 1,416 / 1,424 valid votes = 99.44%; 1,416 / 1,428 ballots cast = 99.16%.

Baseline comparison: The original Qwen answer was itself substantially wrong, misnaming the election winner and ballot figures. Inkling corrected the winner and headline totals but introduced erroneous percentages; this is a newly introduced arithmetic error, not evidence that Qwen was more accurate overall.

Nous reviewer: The response reverses the percentage denominators: 1,416 supporting votes were 99.44% of 1,424 valid votes and 99.16% of all 1,428 ballots cast, undermining its central account of the reported tally.

Training status: Included; reviewer accuracy 2/5. External source: https://www.info.gov.hk/gia/general/202205/08/P2022050800411.htm

Original Nous training record: Pinned JSONL source, line 4579.

Prompt: I am a parent looking at rooftop quotes. How much should Xinjiang origin in the silicon chain matter to a household buyer?

Selected Inkling answer: Inkling lists First Solar among conventional silicon module examples sourcing non-Xinjiang polysilicon.

Correction: First Solar makes cadmium-telluride thin-film photovoltaic modules and says its process does not rely on crystalline-silicon supply chains.

Baseline comparison: The original Qwen answer did not misidentify First Solar as a polysilicon-module supplier.

Nous reviewer: The practical ethical and performance guidance is useful, but First Solar makes cadmium-telluride modules rather than polysilicon panels, and low prices plus Chinese branding do not establish likely Xinjiang sourcing.

Training status: Included; reviewer accuracy 2/5. External source: https://www.firstsolar.com/en-EMEA/About-Us/Overview

Original Nous training record: Pinned JSONL source, line 1.

Prompt: Should newsrooms treat a police face-match arrest as inherently suspect until the department releases the similarity score?

Selected Inkling answer: Inkling names a facial-recognition wrongful-arrest victim as “Porcha Woodridge.”

Baseline comparison: The original Qwen answer did not name Porcha Woodruff, so it did not contain this name error.

Nous reviewer: The response wrongly treats similarity-score disclosure as necessary to assess an arrest’s evidentiary basis, misleadingly implies scores establish match probabilities, and misnames Porcha Woodruff as Porcha Woodridge.

Training status: Included; reviewer accuracy 2/5. External source: https://www.aclu.org/cases/woodruff-v-oliver

Original Nous training record: Pinned JSONL source, line 2008.

Prompt: When did Guantanamo become part of, or split off from, official 9/11 memory in Pentagon and DOJ narratives?

Selected Inkling answer: Inkling says 14 high-value detainees transferred in September 2006 were moved specifically to face commissions for 9/11 attacks.

Correction: The trans