trainable agents
Basic Information
This repository is the official codebase and resource release for the EMNLP 2023 paper 'Character-LLM: A Trainable Agent for Role-Playing'. It provides an end-to-end pipeline to create, train, and run role-specific, trainable language agents that emulate historical or fictional characters using a proposed Experience Reconstruction data generation process. The repo includes training datasets, scripts to generate and parse scene and dialogue data with external LLMs, instructions to assemble model weights from published delta checkpoints, training recipes based on FastChat, and inference server examples for serving models as chatbots. Nine prebuilt character models and associated dataset statistics are provided. The materials are intended for academic research and reproduce experiments described in the paper. Documentation describes dataset formats, data generation steps, SFT training commands, inference server startup, and example single- and multi-turn interview scripts for qualitative evaluation.
Links
Categorization
App Details
-
$$$$ -
$$$$