Hasty Briefsbeta

Jasmine: A Simple, Performant and Scalable Jax-Based World Modeling Codebase

10 days ago
  • #JAX
  • #machine learning
  • #world modeling
  • Jasmine is a JAX-based world modeling codebase designed for performance and scalability.
  • It enables faster reproduction of case studies like CoinRun compared to previous implementations.
  • Jasmine supports fully reproducible training and diverse sharding configurations.
  • The codebase is optimized for data loading, training, and checkpointing.
  • It pairs with curated datasets to establish benchmarking pipelines for model families and architectural ablations.