This paper introduces an LLM-based solution for interactive drama, proposing a prototype drama script and a comprehensive training framework for drama LLMs. The key contributions include the definition of six essential elements for interactive drama—plot, character, thought, diction, spectacle, and interaction—and the development of a narrative chain to enable finer control over the narrative progression. The paper also proposes Auto-Drama, a data generation technique to automatically create drama scripts from arbitrary stories, and Sparse Instruction Tuning (SIT), an enhanced instruction tuning method to train drama LLMs. A multi-aspect evaluation framework is introduced to assess the performance of drama LLMs across five dimensions: scenery, narration, coherency, guidance, and transition. The paper presents a detailed analysis of the challenges in training drama LLMs, including the need for large amounts of drama scripts, the complexity of narrative progression, and the difficulty of following intricate instructions. The proposed solutions address these challenges through the use of narrative chains, which divide the narrative into smaller segments to facilitate smoother and more coherent storytelling. Auto-Drama is used to generate a large number of drama scripts, while SIT ensures that the LLMs can follow complex instructions effectively. The paper also evaluates the performance of the proposed drama LLMs on three manually-written scripts: a detective story, an adventure story, and a classical drama. The results show that the drama LLMs can effectively engage in dialogue with players, generate fluent and rich narratives, and accurately handle plot progression. The evaluation highlights the effectiveness of the proposed methods in enhancing the performance of drama LLMs. The paper also discusses the limitations of the current approach, including the limited modalities supported by the drama LLMs and the complexity of action interactions. The ethical considerations of using drama LLMs are also addressed, emphasizing the importance of responsible use and the need for further research to improve the performance and usability of these systems.This paper introduces an LLM-based solution for interactive drama, proposing a prototype drama script and a comprehensive training framework for drama LLMs. The key contributions include the definition of six essential elements for interactive drama—plot, character, thought, diction, spectacle, and interaction—and the development of a narrative chain to enable finer control over the narrative progression. The paper also proposes Auto-Drama, a data generation technique to automatically create drama scripts from arbitrary stories, and Sparse Instruction Tuning (SIT), an enhanced instruction tuning method to train drama LLMs. A multi-aspect evaluation framework is introduced to assess the performance of drama LLMs across five dimensions: scenery, narration, coherency, guidance, and transition. The paper presents a detailed analysis of the challenges in training drama LLMs, including the need for large amounts of drama scripts, the complexity of narrative progression, and the difficulty of following intricate instructions. The proposed solutions address these challenges through the use of narrative chains, which divide the narrative into smaller segments to facilitate smoother and more coherent storytelling. Auto-Drama is used to generate a large number of drama scripts, while SIT ensures that the LLMs can follow complex instructions effectively. The paper also evaluates the performance of the proposed drama LLMs on three manually-written scripts: a detective story, an adventure story, and a classical drama. The results show that the drama LLMs can effectively engage in dialogue with players, generate fluent and rich narratives, and accurately handle plot progression. The evaluation highlights the effectiveness of the proposed methods in enhancing the performance of drama LLMs. The paper also discusses the limitations of the current approach, including the limited modalities supported by the drama LLMs and the complexity of action interactions. The ethical considerations of using drama LLMs are also addressed, emphasizing the importance of responsible use and the need for further research to improve the performance and usability of these systems.