TRL v1.0: Post-Training Library Built to Move with the Field

TRL v1.0 marks the evolution of a research tool into a dependable, production-ready library for post-training large language models. It is specifically designed to thrive in the constantly changing field of AI alignment, where core methods and assumptions are frequently redefined. To provide stability amidst this chaos, the library introduces a unique dual structure with a stable core for proven methods and a separate experimental layer for new, fast-moving algorithms. This "chaos-adaptive" approach intentionally limits deep abstractions, allowing the library to adapt to new paradigms without breaking the many downstream projects that depend on it.