<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Arete on Local First AI</title><link>https://localfirstai.eu/tags/arete/</link><description>Recent content in Arete on Local First AI</description><image><title>Local First AI</title><url>https://localfirstai.eu/images/og-default.png</url><link>https://localfirstai.eu/images/og-default.png</link></image><generator>Hugo</generator><language>en-us</language><lastBuildDate>Sun, 08 Nov 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://localfirstai.eu/tags/arete/index.xml" rel="self" type="application/rss+xml"/><item><title>We Measured a Values Fine-Tune. It Removes the Behaviour, Adds No Knowledge, and Costs Nothing on Our Translations.</title><link>https://localfirstai.eu/posts/2026-11-08-we-measured-a-values-fine-tune/</link><pubDate>Sun, 08 Nov 2026 00:00:00 +0000</pubDate><guid>https://localfirstai.eu/posts/2026-11-08-we-measured-a-values-fine-tune/</guid><description>Nous Research&amp;#39;s Arete-27B, Qwen3.8-27B post-trained to drop Chinese state-aligned answers, against its base on a Mac mini, both Q4_K_M, graded blind. Three things proved: the fine-tune does what it says and survives the community quantisation (any propaganda 24.3 % → 1.4 % on 70 of the vendor&amp;#39;s prompts, their judge prompt); it costs nothing on 60 Soviet-era translation units it was never trained for; it adds no knowledge, it removes a denial template.</description></item></channel></rss>