Agent-FLAN

Agent-FLAN旨在为大型语言模型提供有效的Agent Tuning数据和方法。它包含AgentInstruct、Toolbench以及定制的负样本数据集，用于Agent任务和工具利用的微调。该仓库提供了一系列数据集，包括agent_instruct_react、agent_instruct_tflan、toolbench_instruct_j1s1_3k等多个split，数据以JSONL格式存储，包含对话（角色、内容、损失）和ID等特征。Agent-FLAN基于Llama2-chat的对话格式，通过混合训练AgentInstruct、ToolBench和ShareGPT数据集生成。本仓库使用Apache 2.0协议授权。

Agent-FLAN aims to provide effective Agent Tuning data and methods for large language models. It includes AgentInstruct, Toolbench, and customized negative sample datasets for fine-tuning on Agent tasks and tool utilization. This repository offers a collection of datasets with multiple splits, such as agent_instruct_react, agent_instruct_tflan, and toolbench_instruct_j1s1_3k. All data is stored in JSONL format, containing fields like dialogues (with role, content, and loss information) and IDs. Agent-FLAN is built upon the dialogue format of Llama2-chat, and is generated via joint training on the AgentInstruct, ToolBench, and ShareGPT datasets. This repository is licensed under the Apache 2.0 license.