radixark/miles

Train frontier models with reinforcement learning at scale! Miles is an open-source framework for large-scale RL post-training across LLMs and VLMs. At this scale, RL is not just about choosing an…

Miles v0.1 开源了：一套现成的"大模型训练流水线"，让企业自己训专属模型这件事，从烧钱赌博变成工程问题，属于SGLang 系那帮人做的开源 RL 后训练框架。 以前想让模型学会干你公司的活（写代码、操作软件、处理客服），要么找大厂定制，要么自己养一支底层团队从零搭，Miles 把这套东西打包好了，Docker 一键起，换上你的数据和任务就能训。…

Miles v0.1: Production-Level Post-Training A full-stack RL system for frontier-scale post-training, built with SGLang for rollout and Megatron-LM for training, now open-sourced.

Preregistered audit of 52,988 LLM-judge requests: same candidates re-ranked at 0.40 Spearman where 0.90 was required. Next-day byte-identical replays hit 0.78 vs a required 0.99. Paper: Our read +…

Today we're launching Miles v0.1, an open-source RL framework for LLMs and multimodal models. RL training is easy to start and hard to debug. Miles helps you ensure your run is correct, use hardware…