Hugging Face· PlatformDeveloper tool

PipelineRL

ServiceNow and Hugging Face open-source PipelineRL, an experimental reinforcement learning implementation that addresses the trade-off between inference throughput and on-policy data collection by using inflight weight updates during RL training. It enables high inference throughput while keeping data on-policy, achieving competitive results on reasoning benchmarks compared to Open-Reasoner-Zero when training 7B and 32B models.

From the source

We are excited to open-source PipelineRL, an experimental RL implementation that tackles a fundamental challenge in large-scale Reinforcement Learning with LLMs: the trade-off between inference throughput and on-policy data collection.

huggingface.co

Announced

25 Apr 2025 22:37 UTC

Detected

03 Aug 2026 21:12 UTC (+464d after)