--- title: Smaller model, sharper insights: How Atos approached... description: A fine-tuned 1-billion-class model beat GPT-4.1 on a structured extraction task for a German public sector client, using Red Hat OpenShift AI and just 3 NVIDIA... canonical: https://daily.dev/posts/smaller-model-sharper-insights-how-atos-approached-structured-outputs-drvpznghd twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Smaller model, sharper insights: How Atos approached structured outputs | daily.dev og:description: A fine-tuned 1-billion-class model beat GPT-4.1 on a structured extraction task for a German public sector client, using Red Hat OpenShift AI and just 3 NVIDIA... og:url: https://daily.dev/posts/smaller-model-sharper-insights-how-atos-approached-structured-outputs-drvpznghd og:image: https://api.daily.dev/og/posts/drvpZNGhd.png og:image:alt: Smaller model, sharper insights: How Atos approached structured outputs og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Smaller model, sharper insights: How Atos approached structured outputs **[Red Hat Developer](https://daily.dev/sources/rhdev)** · 10 min read · 0 upvotes · 0 comments ## Summary A fine-tuned 1-billion-class model beat GPT-4.1 on a structured extraction task for a German public sector client, using Red Hat OpenShift AI and just 3 NVIDIA L40 GPUs. Atos fine-tuned Phi-4 Mini 3.8B, Qwen 2.5 1.5B, and Qwen 2.5 7B as student models on synthetic data generated by GPT-4.1 as a teacher model. JSON validity for Phi-4 Mini jumped from 1% before tuning to 99% after. On a custom 34-point rubric, fine-tuned Qwen 2.5 7B scored 30.8 points, beating GPT-4.1 with a detailed prompt at 30.5 points. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Can a fine-tuned small model actually outperform GPT-4.1 on a real task? Yes. In a proof of value for a German public sector client, a fine-tuned Qwen 2.5 7B model scored 30.8 points on a custom 34-point structured extraction rubric, beating GPT-4.1's best detailed-prompt score of 30.5. Fine-tuned Phi-4 Mini 3.8B also beat GPT-4.1's short-prompt score, reaching about 30 points versus roughly 29. _Teams choosing between frontier APIs and fine-tuned small models can track real-world benchmarks like this on daily.dev._ ### How much GPU hardware do I need to fine-tune a small language model for a custom task? Only 3 NVIDIA L40 GPUs were needed to fine-tune three student models (Phi-4 Mini 3.8B, Qwen 2.5 1.5B, and Qwen 2.5 7B) across 27 variations using supervised fine-tuning and orthogonal subspace fine-tuning on Red Hat OpenShift AI, with no GPU required for the synthetic data generation step itself. _Anyone scoping an on-premises fine-tuning project can keep up with hardware benchmarks like this via daily.dev._ ### How do you improve JSON output reliability in a fine-tuned LLM for structured extraction? Fine-tuning on domain-specific synthetic data dramatically improves JSON validity: the Phi-4 Mini base model produced valid JSON only 1% of the time, but after a single epoch of fine-tuning that jumped to 92%, reaching about 99% with further training, matching or beating GPT-4.1's 96% JSON validity rate. _Developers chasing reliable structured output from LLMs can follow fine-tuning results like these on daily.dev._ ## Similar posts on daily.dev - [I ran 14b and 20b local AI models on my laptop for the same tasks, and result was surprising](https://daily.dev/posts/i-ran-14b-and-20b-local-ai-models-on-my-laptop-for-the-same-tasks-and-result-was-surprising-vqclmy07p) · XDA Developers · 0 upvotes · 0 comments - [The rise of small language models and why open source is winning](https://daily.dev/posts/the-rise-of-small-language-models-and-why-open-source-is-winning-laovm7ruy) · All Things Open · 38 upvotes · 0 comments - [The Agentic Analytics Benchmark, round two: GPT-6, Claude Opus 5.5, Claude Sonnet 5.5](https://daily.dev/posts/the-agentic-analytics-benchmark-round-two-gpt-6-claude-opus-5-5-claude-sonnet-5-5-l8whcfkkx) · ClickHouse · 1 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#deep-learning](https://daily.dev/tags/deep-learning), [#red-hat](https://daily.dev/tags/red-hat) [View this post on daily.dev](https://daily.dev/posts/smaller-model-sharper-insights-how-atos-approached-structured-outputs-drvpznghd) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Smaller model, sharper insights: How Atos approached structured outputs","url":"https://daily.dev/posts/smaller-model-sharper-insights-how-atos-approached-structured-outputs-drvpznghd","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/smaller-model-sharper-insights-how-atos-approached-structured-outputs-drvpznghd"},"datePublished":"2026-10-27T13:37:40.428Z","dateModified":"2026-10-27T13:40:10.715Z","description":"A fine-tuned 1-billion-class model beat GPT-4.1 on a structured extraction task for a German public sector client, using Red Hat OpenShift AI and just 3 NVIDIA...","image":"https://media.daily.dev/image/upload/s--CxzD6vbw--/f_auto/v1722860399/public/Placeholder%2005","thumbnailUrl":"https://media.daily.dev/image/upload/s--CxzD6vbw--/f_auto/v1722860399/public/Placeholder%2005","isAccessibleForFree":true,"articleSection":"Red Hat Developer","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Red Hat Developer","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/2a8895f3216c4900bdc455585a4ad3e4","url":"https://daily.dev/sources/rhdev"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/smaller-model-sharper-insights-how-atos-approached-structured-outputs-drvpznghd","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,deep-learning,red-hat","timeRequired":"PT10M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Red Hat Developer","item":"https://daily.dev/sources/rhdev"},{"@type":"ListItem","position":3,"name":"Smaller model, sharper insights: How Atos approached structured outputs"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/smaller-model-sharper-insights-how-atos-approached-structured-outputs-drvpznghd#faq","mainEntity":[{"@type":"Question","name":"Can a fine-tuned small model actually outperform GPT-4.1 on a real task?","acceptedAnswer":{"@type":"Answer","text":"Yes. In a proof of value for a German public sector client, a fine-tuned Qwen 2.5 7B model scored 30.8 points on a custom 34-point structured extraction rubric, beating GPT-4.1's best detailed-prompt score of 30.5. Fine-tuned Phi-4 Mini 3.8B also beat GPT-4.1's short-prompt score, reaching about 30 points versus roughly 29. Teams choosing between frontier APIs and fine-tuned small models can track real-world benchmarks like this on daily.dev."}},{"@type":"Question","name":"How much GPU hardware do I need to fine-tune a small language model for a custom task?","acceptedAnswer":{"@type":"Answer","text":"Only 3 NVIDIA L40 GPUs were needed to fine-tune three student models (Phi-4 Mini 3.8B, Qwen 2.5 1.5B, and Qwen 2.5 7B) across 27 variations using supervised fine-tuning and orthogonal subspace fine-tuning on Red Hat OpenShift AI, with no GPU required for the synthetic data generation step itself. Anyone scoping an on-premises fine-tuning project can keep up with hardware benchmarks like this via daily.dev."}},{"@type":"Question","name":"How do you improve JSON output reliability in a fine-tuned LLM for structured extraction?","acceptedAnswer":{"@type":"Answer","text":"Fine-tuning on domain-specific synthetic data dramatically improves JSON validity: the Phi-4 Mini base model produced valid JSON only 1% of the time, but after a single epoch of fine-tuning that jumped to 92%, reaching about 99% with further training, matching or beating GPT-4.1's 96% JSON validity rate. Developers chasing reliable structured output from LLMs can follow fine-tuning results like these on daily.dev."}}]} ```