AgentOhana gives heterogeneous agent data a common training format

Interactive environments record an agent's work in different ways. A web task may produce page observations, while another environment returns structured tool results. AgentOhana develops a unified representation and training pipeline for these heterogeneous trajectories.

Preserve the interaction in a shared format

The data pipeline aggregates agent trajectories from distinct environments and standardizes them into a consistent structure. That representation carries the sequence of observations and actions into a common data loader.

A shared format gives training code a regular interface to varied tasks. The environment-specific content remains part of the example, while the surrounding representation makes the trajectory usable within the same training workflow.

Control how sources enter training

AgentOhana addresses the balance between data sources as examples are loaded for distributed training. It also preserves independent randomness across devices during data partitioning and model training.

These details connect dataset preparation with the actual stream of examples seen by the model. Source balance and sampling become explicit parts of the training system that researchers can inspect.

Evaluate learning across environments

The work introduces xLAM-v0.1 as an action model trained through the pipeline and evaluates it on agent benchmarks. The release connects a data-standardization method, a training implementation, and a model artifact.

The contribution is useful wherever a training collection combines multiple kinds of interaction. A common representation helps researchers examine the examples together; explicit source handling helps them understand the mixture used by a run. Those two pieces provide a concrete starting point for studying how environment diversity affects learned agent behavior.

Jianguo Zhang, Tian Lan, and collaborators developed AgentOhana at Salesforce Research, with Weiran Yao among the authors. The report was first published in 2024. Read the paper and explore the associated action-model resources.