metadata
license: apache-2.0
Video-Coder Training Data
This repository contains the training data for Video-Coder: Reinforcing Agentic Coding for Video Generation.
🌐 Project Page | 📖 Paper | 💻 Code
👀 Intro
Video-Coder is a code-driven agent for video generation. It iteratively writes, executes, and refines Blender Python programs to construct dynamic 3D scenes and render draft videos, providing structural and spatial guidance for downstream video generators.
We design a dedicated data pipeline and curate two high-quality datasets, Video-Coder-SFT-10k and Video-Coder-RL-4k, which contain diverse prompts and ground-truth reference videos.
Dataset Files
The Video-Coder training data consists of JSON files tailored for different training stages:
sft_data.json: Used for Supervised Fine-Tuning (SFT) cold start.rl_data.json: Used for Reinforcement Learning (RL) training using GRPO.
🎥 Demo
Citation
If you find this work helpful for your research, please consider citing: