Research
ActionStudio and the data pipeline behind action-model training
A unified trajectory format, data verification, and distributed training support heterogeneous agent datasets.
Agent datasets record conversations, actions, and observations in different formats. ActionStudio creates a common path from those heterogeneous records to a training run for large action models.
Give trajectories a shared representation
The framework introduces Unified Format 2.0 to represent agent trajectories from different environments. A common format lets preprocessing and training code work with a consistent structure while preserving the sequence of the interaction.
ActionStudio also includes a critique-and-filter pipeline for preparing examples. These steps connect the quality of an individual trajectory with the collection that reaches the training process.
Inspect data as it changes form
A real-time verifier presents examples before and after processing stages, including conversational templating and encoding. This makes the transformation from stored trajectory to model input visible during development.
The training handler supports supervised fine-tuning and preference-based learning, with configurable training setups and multi-node execution. Data processing and distributed training sit within the same framework, giving researchers a shared implementation for both stages.
Connect the release with the training system
The work releases ActionStudio and the actionstudio-98k collection of curated trajectories. Its experiments evaluate training throughput and the behavior of models trained with the framework on agent benchmarks.
The contribution is an inspectable data-to-training path. A developer can examine the original action sequence, its normalized representation, the tokens presented to the model, and the settings used to train it. That continuity is useful when investigating how a data-processing choice affects the behavior learned from an interaction.
Jianguo Zhang, Thai Hoang, and colleagues developed ActionStudio at Salesforce AI Research. Haolin Chen and Weiran Yao contributed to the work, which appeared at EMNLP 2025. Read the paper and explore the associated action-model resources.