Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Sep 7·5 min read
Iris: Climbing to the Search Frontier
TL;DR
Iris-mini and Iris-pro are advanced search agents trained on large-scale datasets, achieving top performance in open-source benchmarks. They utilize a novel training methodology that combines supervised fine-tuning and reinforcement learning to optimize search capabilities.
✦ Why It Matters
Explore the released model weights and training recipes to enhance your own search applications.
Key Takeaways
- Utilize the Iris models for advanced search capabilities in your applications.
- Implement context management strategies to improve inference performance.
- Explore the training recipes for insights into effective model training.
- Stay updated on the release of model weights for practical experimentation.
How It Works
Iris agents are trained using a method that combines supervised fine-tuning and reinforcement learning. Tasks are generated from web hyperlinks, creating complex question chains that require reasoning.
The SFT-RL climbing approach alternates between refining the model through supervised learning and optimizing it via reinforcement learning, allowing the agents to learn from both structured data and real-time search interactions.