Jack of All Trades (JAT) generalist Transformer agent with JAT dataset and GPT-Neo encoder

AI Impact Summary

Jack of All Trades (JAT) extends the generalist-agent concept by releasing a GPT-Neo–based Transformer capable of multi-modal perception and sequential decision-making, building on the Gato approach. The package includes the JAT dataset and a spectrum of expert RL agents across environments like Atari, BabyAI, Meta-World, and MuJoCo, with the model and dataset exposed via the HuggingFace Hub. This enables cross-domain benchmarking and rapid prototyping of multi-modal agents, though teams should evaluate dataset coverage, environment fidelity, and production-readiness before deploying in robotics or real-time decision tasks.