ToolAlpaca

ToolAlpaca is a tool-learning dataset released with the paper "ToolAlpaca: Generalized Tool Learning for Language Models with 3000 Simulated Cases" (Tang et al., 2023). The published corpus contains 3,938 tool-use instances drawn from more than 400 real-world tool APIs spanning 50 distinct categories (the repo's data/train_data.json actually holds 4,255 instances across 468 tool records). Each tool record includes an LLM-generated OpenAPI specification, natural-language documentation, and a set of instances; each instance is a ReAct-style trajectory with a user input, a sequence of intermediate_steps (each a [Action, Action Input, Thought] triple paired with a tool Observation), a Final Thought, and a final natural-language output. The data was produced automatically in a multi-agent simulation environment (a user agent issues instructions, an assistant agent decides actions, and a simulator agent plays the tool executor), with no human-written trajectories. Two held-out evaluation splits are provided: eval_simulated.json (10 simulated APIs) and eval_real.json (11 real APIs, some requiring authentication), each supplying Instructions and Golden_Answers (reference action sequences).