**AgentOhana: Design Unified Data and Training Pipeline for Effective Agent Learning** This paper introduces AgentOhana, a comprehensive solution to address the challenges of using large language models (LLMs) for agent-based tasks. AgentOhana aggregates and standardizes multi-turn trajectories from diverse environments, streamlining the creation of a generic data loader optimized for agent training. The training pipeline maintains equilibrium across different data sources and preserves independent randomness during dataset partitioning and model training. Additionally, the paper presents xLAM-v0.1, a large action model tailored for AI agents, which demonstrates strong performance across various benchmarks. **Key Contributions:** 1. **Innovative Solution to Data Heterogeneity:** AgentOhana addresses the challenges of consolidating heterogeneous data sources for multi-turn LLM agent trajectories. 2. **Extensive Environmental Coverage:** AgentOhana incorporates data from ten distinct environments, spanning a wide array of scenarios. 3. **Data Standardization and Unification:** A systematic approach to standardize and unify LLM agent data into a consistent format, enabling a generic data loader. 4. **Large Agent Model:** xLAM-v0.1, a robust large action model, showcases exceptional performance across rigorous benchmarks. **Methodology:** - **Homogeneous Multi-turn Agent Trajectory Standardization:** A unified JSON dictionary format captures all relevant content of each trajectory, including user queries, model names, and scores. - **AgentRater:** A method to rate agent trajectories based on public or close-world models, ensuring high-quality trajectories. - **Generic Dataloader:** Facilitates seamless integration of diverse datasets into the training process, ensuring randomness and reproducibility. **Experiments:** - **Training:** Supervised fine-tuning of xLAM-v0.1 on the Mixtral-8x7B-Instruct-v0.1 model, conducted on 8 Nvidia H100 GPUs using the QLoRA framework. - **Benchmarks:** Evaluations on Webshop, HotpotQA, ToolEval, and MINT-Bench, demonstrating superior performance compared to other models. **Conclusion:** AgentOhana represents a significant advancement in handling diverse data for multi-turn LLM agent trajectories, providing a robust framework for researchers and practitioners to enhance the capabilities of autonomous agents powered by LLMs.**AgentOhana: Design Unified Data and Training Pipeline for Effective Agent Learning** This paper introduces AgentOhana, a comprehensive solution to address the challenges of using large language models (LLMs) for agent-based tasks. AgentOhana aggregates and standardizes multi-turn trajectories from diverse environments, streamlining the creation of a generic data loader optimized for agent training. The training pipeline maintains equilibrium across different data sources and preserves independent randomness during dataset partitioning and model training. Additionally, the paper presents xLAM-v0.1, a large action model tailored for AI agents, which demonstrates strong performance across various benchmarks. **Key Contributions:** 1. **Innovative Solution to Data Heterogeneity:** AgentOhana addresses the challenges of consolidating heterogeneous data sources for multi-turn LLM agent trajectories. 2. **Extensive Environmental Coverage:** AgentOhana incorporates data from ten distinct environments, spanning a wide array of scenarios. 3. **Data Standardization and Unification:** A systematic approach to standardize and unify LLM agent data into a consistent format, enabling a generic data loader. 4. **Large Agent Model:** xLAM-v0.1, a robust large action model, showcases exceptional performance across rigorous benchmarks. **Methodology:** - **Homogeneous Multi-turn Agent Trajectory Standardization:** A unified JSON dictionary format captures all relevant content of each trajectory, including user queries, model names, and scores. - **AgentRater:** A method to rate agent trajectories based on public or close-world models, ensuring high-quality trajectories. - **Generic Dataloader:** Facilitates seamless integration of diverse datasets into the training process, ensuring randomness and reproducibility. **Experiments:** - **Training:** Supervised fine-tuning of xLAM-v0.1 on the Mixtral-8x7B-Instruct-v0.1 model, conducted on 8 Nvidia H100 GPUs using the QLoRA framework. - **Benchmarks:** Evaluations on Webshop, HotpotQA, ToolEval, and MINT-Bench, demonstrating superior performance compared to other models. **Conclusion:** AgentOhana represents a significant advancement in handling diverse data for multi-turn LLM agent trajectories, providing a robust framework for researchers and practitioners to enhance the capabilities of autonomous agents powered by LLMs.