<?xml version="1.0" encoding="utf-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
        <title>Fine-Tuned Qwen 0.8B Beat Prompted GPT-5.4 on Preply's Grammar Test (Company Preprint)</title>
        <link>https://peertube.watch/videos/watch/5c295110-8ea2-4f5a-be63-090aacb50a1b</link>
        <description>Peace Call. Mr. Peace is an AI-made character; voice is AI-generated; script written with AI assistance. No affiliate links in this video. How this site is run: https://getpeacecall.com/verify/?utm_source=peertube&amp;utm_medium=video&amp;utm_campaign=underdog-preply According to a company-authored preprint by Preply about its own product, a fine-tuned Qwen3.5-0.8B scored 82.74% precision and 85.16% recall at the strictest level (concept + exact span + correctness) on Preply's 689-utterance Golden test, vs 67.81% / 62.87% for prompted GPT-5.4. Untuned, the same 0.8B model scored 0.01% / 0.19% - fine-tuning made the difference in this test. GPT-5.6 Sol (prompted) scored 74.37% / 76.04%. Cost: modeled recurring detection cost was about 6.2% of the GPT-5.4 baseline. Training, deployment, upstream processing, and fallback calls were excluded. A feature-level online test showed +15.8% learner engagement, but that measures the whole feature, not the model swap alone. Source: Celikik, Peleteiro Ramallo, Morales (Preply), 'Grammar Concept Annotation at Scale: Deployed Fine-Tuned Small Language Models Outperform Prompted Frontier Models', arXiv:2610.10827 (v1 Oct 7, 2026; v2 Oct 11, 2026) - PREPRINT, company-authored, not yet peer reviewed. https://arxiv.org/abs/2610.10827 . Numbers from Table 2, Section 7 and Appendix D.2. The training data was generated/reviewed by GPT-5.4; sizes used different recipes (not a controlled comparison). Names are trademarks of their respective owners; Peace Call is not affiliated with or endorsed by them. Charts drawn by code from the source's numbers. Paper: https://arxiv.org/abs/2610.10827 (version v2, posted Oct 11, 2026; v1 Oct 7, 2026). AI Disclosure: Mr. Peace is a fictional AI character using a synthetic voice. This video is AI-assisted and does not depict a real person speaking. Research Disclosure: Based on a research preprint that has not been peer-reviewed. Results apply to the reported benchmarks and experimental conditions, not overall model superiority. Trademark Disclosure: Names are trademarks of their respective owners. Peace Call is not affiliated with or endorsed by them.</description>
        <lastBuildDate>Wed, 04 Nov 2026 08:05:04 GMT</lastBuildDate>
        <docs>https://validator.w3.org/feed/docs/rss2.html</docs>
        <generator>PeerTube - https://peertube.watch</generator>
        <image>
            <title>Fine-Tuned Qwen 0.8B Beat Prompted GPT-5.4 on Preply's Grammar Test (Company Preprint)</title>
            <url>https://peertube.watch/lazy-static/avatars/3b284627-b75f-409e-ac2e-10f4c2946f43.png</url>
            <link>https://peertube.watch/videos/watch/5c295110-8ea2-4f5a-be63-090aacb50a1b</link>
        </image>
        <copyright>All rights reserved, unless otherwise specified in the terms specified at https://peertube.watch/about and potential licenses granted by each content's rightholder.</copyright>
        <atom:link href="https://peertube.watch/feeds/video-comments.xml?videoId=5c295110-8ea2-4f5a-be63-090aacb50a1b" rel="self" type="application/rss+xml"/>
    </channel>
</rss>