<?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>Dpo on GPUburnout | Jun Park</title><link>https://gpuburnout.com/tags/dpo/</link><description>Recent content in Dpo on GPUburnout | Jun Park</description><image><title>GPUburnout | Jun Park</title><url>https://gpuburnout.com/images/og-default.png</url><link>https://gpuburnout.com/images/og-default.png</link></image><generator>Hugo -- 0.155.2</generator><language>en-us</language><lastBuildDate>Sun, 29 Mar 2026 00:00:00 +0000</lastBuildDate><atom:link href="https://gpuburnout.com/tags/dpo/index.xml" rel="self" type="application/rss+xml"/><item><title>7 Out of 8 - How DPO Finally Worked</title><link>https://gpuburnout.com/posts/s4-ch3-dpo-finally-worked/</link><pubDate>Sun, 29 Mar 2026 00:00:00 +0000</pubDate><guid>https://gpuburnout.com/posts/s4-ch3-dpo-finally-worked/</guid><description>In which the same technique that failed nine times on the 1B succeeds on the 2B - and the reason why is the whole point of Season 4.</description></item><item><title>Nine Experiments, Nine Funerals</title><link>https://gpuburnout.com/posts/s3-ch3-garbage-survived-finetuning/</link><pubDate>Sat, 21 Mar 2026 00:00:00 +0000</pubDate><guid>https://gpuburnout.com/posts/s3-ch3-garbage-survived-finetuning/</guid><description>A controlled experiment on why post-training alignment can&amp;#39;t fix pretraining contamination - and what the data proves.</description></item></channel></rss>