TRL v1.0: Post-Training Library stabilizes for production

AI Impact Summary

TRL v1.0 shifts TRL from a research codebase to production-grade infrastructure, expanding to 75+ post-training methods. The library now exposes a stable core with semantic versioning and an experimental surface that can move faster, allowing production teams to rely on trainers like SFTTrainer and DPOTrainer while evaluating new methods via experimental APIs. Downstream projects that tightly depended on legacy contracts (e.g., Unsloth, Axolotl) will feel breaking changes unless they follow the migration guide, but the split helps minimize disruption by isolating unstable APIs. With TRL downloaded around 3 million times per month, adoption is broad, so teams should plan upgrade paths and leverage the dedicated migration guide to minimize runtime incidents.