Industry11月14日 02:18

Anyscale 举办 RL 与数据 meetup，Mercor、Coinbase 分享实战经验

Anyscale@anyscalecompute

Post

Why we picked it

这场meetup把工业界RL训练的实战细节凑齐了：Mercor用SkyRL训练，Coinbase在造领域专家模型，还有1M-token上下文训练的坑。做RL后训练或用Ray的工程团队，这些一手经验比论文实在。

Anyscale 昨晚举办了一场满场的强化学习（RL）与数据主题 meetup，共六位讲者分享生产环境经验。Mercor 的 Tao Tianhua 讲解如何用 SkyRL 做 RL 训练，Anyscale 的 Charlie Ruan 介绍了 SkyRL 的 day-zero 模型支持与新功能，Anyscale 与 Lila Sciences 则展示了 1M-token 上下文训练的工作。

Body

技术细节上，Xinyu 讲解了基于 Ray 构建的原生 sandbox 环境，Coinbase 的 Yao Ma 分享了训练自定义专家模型的实践，Alex G. Shaw 介绍了 RL 环境设计与 Harbor 框架。整场围绕 SkyRL、Ray、Harbor 这条 RL 训练工具链展开，覆盖从模型支持到环境搭建的完整链路。

对做 RL 后训练的工程团队来说，这次分享提供了多个真实企业案例——Mercor 用 SkyRL 训练模型、Coinbase 构建领域专家模型——比纯研究论文更接近生产落地。使用 Ray 分布式框架或考虑 SkyRL 的开发者可以从这些 day-zero 支持经验中直接借鉴。

Original · Anyscale

Completely packed RL & data meetup last night! - @TaoTianhua spoke about RL at @mercor with SkyRL - @charlie_ruan shared day-zero model support, the latest SkyRL features - @avigyabasnet talked about 1M-token context training work from @anyscalecompute and @LilaSciences - @xinyzng spoke about native @raydistributed sandboxes - @yaoma_ai shared @coinbase’s work on building custom specialist models - @alexgshaw talked about RL environments and @harborframework

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Industry" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include SkyRL, Ray, Anyscale, reinforcement learning.

Source attribution: curated from Anyscale (@anyscalecompute).

#SkyRL #Ray #Anyscale #reinforcement learning #RL environments

Keep reading

Recent stories with the same tag or section

How-to Builders03:03

Chinese contentAnyscale 网络研讨会：强化学习后训练与 SkyRL 实战

Anyscale 要开一场 RL 后训练的实操研讨会，从 GRPO、OPD 蒸馏讲到 SkyRL 的异步训练和 rollout 调优，还会拆解怎么给终端 agent 建 eval 和自定义环境。想自己搭 RL 训练管线的开发者值得报名听听。

#SkyRL #Anyscale #GRPO #reinforcement learning

Anyscale@anyscalecomputeSource

How-to BuildersSelected02:01

Anyscale session on LLM post-training, RLVR and SkyRL

The most useful part is the terminal-agent case study, which shows how evals, custom environments and RL infrastructure fit together in a real post-training pipeline rather than as isolated concepts. Engineers building agents or running reinforcement learning fine-tuning can use SkyRL and the async rollout workflow to structure their own training runs. The OPD and MOPD sections also outline distillation methods rarely covered in tutorials.

#SkyRL #GRPO #RLVR #reinforcement learning

Anyscale@anyscalecomputeSource

How-to BuildersSelected04:03

Anyscale Offers Hands-On Training on LLM Post-Training with SkyRL

Agentic RL with GRPO and on-policy distillation are moving from research papers into standard production toolkits, and this course shows how teams actually wire them up with SkyRL, evals and custom environments. Engineers building agent training pipelines can follow the case-study structure directly. Registration is via Anyscale's official link, so seats and schedule depend on their listed availability.

#Anyscale #SkyRL #GRPO #On-policy Distillation

Event

Anyscale@anyscalecomputeSource

Models BuildersSelected23:29

Reflection-Beam-1.0 gets Day-0 support in vLLM

Day-0 serving and RL support means practitioners can deploy and fine-tune a 501B-parameter sparse model without waiting for ecosystem catch-up. Teams running large-scale inference can start with vLLM and add reinforcement learning through SkyRL's LoRA or MXFP8 paths. The 1M-token context window puts Beam in direct competition with other frontier sparse models for long-document and agentic workloads.

#Reflection-Beam-1.0 #vLLM #SkyRL #reasoning models

Official accountvLLM@vllm_projectSource

Models BuildersSelected21:28

Reflection-Beam-1.0 Gains Day-0 Support in vLLM

The sparse-activation design at 501B/23B active with a 1M-token window makes Beam a candidate for long-context workloads without full MoE serving costs. Teams running vLLM can deploy it immediately and fine-tune with SkyRL's LoRA or MXFP8 RL paths. Compared with other large reasoning models, Beam's Day-0 coverage of both serving and RL tooling is the differentiator.

#Reflection-Beam-1.0 #vLLM #open-source models #SkyRL

Official accountvLLM@vllm_projectSource

How-to BuildersOfficialSelected00:24

AWS shows SkyRL recipe for post-training Qwen3-VL-8B with GRPO

Multimodal RL post-training of open VLMs is becoming a reproducible recipe rather than a research one-off, with the trained LoRA adapter ready to deploy directly on SageMaker. Teams building vision-language applications can follow the steps to fine-tune Qwen3-VL-8B on their own data using GRPO without managing infrastructure manually. It also gives Qwen a concrete deployment path on AWS infrastructure.

#SkyRL #SageMaker HyperPod #Qwen3-VL #GRPO

Primary sourceAWS Machine Learning Blog@Nilesh PSSource