In the news
Comparison of techniques for fine-tuning open-weight models for entity extraction from radiology reports
arXiv cs.AI · Published · 3 min read
In 30 seconds
- What happened
- Researchers showed that fine-tuned open-weight Gemma-3-12B matched GPT-4o performance on extracting intracranial hemorrhage labels from radiology reports using distilled real data.
- Why it matters
- Healthcare engineers building on-premises clinical NLP systems need private, cost-effective alternatives to proprietary models for structured data extraction from medical text.
- Watch out
- Synthetic training data failed to improve performance over the base model at any size; results are specific to one narrow task and may not generalize to other clinical extractions.
- fine-tun
- open-weight
- gpt
The patterns behind this
- Agentic Context Engineering (Evolving Playbook)
- Machine Learning Model-Based Routing
- Context Editing & Tool-Result Clearing
Each one covers how the technique works, when it earns its cost, and where it breaks.
The Agent Architect
One pattern, one tradeoff, one production failure story. A short weekly briefing for people building agentic systems.
Weekly email, one-click unsubscribe. We only use your address to send the briefing.