harness 1

We have hosted the application harness 1 in order to run this application in our online workstations with Wine or directly.

Quick description about harness 1:

Harness-1 is a 20B search agent trained with reinforcement learning inside a stateful retrieval harness. It is designed for long-horizon search tasks where the model must search, inspect documents, curate evidence, verify claims, and decide when enough evidence has been gathered. The harness externalizes search state, including candidate documents, evidence links, verification records, and budget-aware context. This lets the policy focus on higher-level decisions instead of trying to keep every detail inside the model context. The repository includes inference utilities, training scripts, evaluation runners, dataset tools, and documentation for running the released checkpoint. Its main value is showing how a smaller open model can approach advanced search-agent behavior through structured retrieval state and reinforcement learning.

Features:

• 20B reinforcement-trained search agent
• Stateful retrieval harness
• Recoverable evidence and verification records
• vLLM-based local serving support
• BrowseComp+ evaluation workflow
• Training, inference, and ablation scripts

Programming Language: Python.
Categories: