radixarkOfficial·7 Agent Skills Included

miles

Train large AI models with reinforcement learning at scale

Runs reinforcement learning post-training for large language models at trillion-parameter scale. Combines fast SGLang rollout with Megatron-LM training, async pipelines, and low-precision FP8 and FP4 support. Removes manual work around weight updates, fault recovery, and multi-GPU cluster setup. Includes agent skills for CI log analysis, test timing calibration, and safe code refactoring.

2,304|421

↗

npx skills add radixark/miles --all -g -y

Available:

Directs the AI agent to read the repository's code style and launch-script rules before creating or modifying training, tooling, or model files.

All Skills in This Repository (7)

Pure Emerald Level Indicators

📦 In Repo

radixark

setup-ci-host

Prepare a physical CI host filesystem layout for miles GitHub Actions runners.

manage-gh-runners

Provision, list, update, and deregister GitHub Actions self-hosted runners via Docker.

mechanical-refactor-verify

Validate mechanical refactors by running a transform script against a target commit.

ci-e2e-time-tune

Calibrates E2E test time estimates from complete GitHub Actions CI logs.

doc-dev

Keeps source files consistent with their governing header and central documentation.

ci-fetch-log

Fetch complete GitHub Actions job logs and diagnose failed CI jobs from saved evidence.

neon-access

Executes arbitrary PostgreSQL SQL against the Neon database through an authorized GitHub Actions workflow.

Official

Advanced

Frequently Asked Questions

FAQPage Schema

How to install Miles?▼

Run `npx skills add radixark/miles --all -g -y` in your terminal to install all skills in this suite globally.

What is Miles used for?▼

Miles is a reinforcement learning framework for post-training large language models, pairing SGLang for fast rollout with Megatron-LM for scalable training on GPU clusters.

Which models and hardware does Miles support?▼

It supports frontier models like DeepSeek-V4, Kimi-K3, GLM-5.2, and Qwen3.5 on NVIDIA H100 through GB300 and AMD MI300X through MI355X GPUs.

Can AI agents help manage Miles CI and refactoring?▼

Yes. The bundled skills let an agent fetch and diagnose CI logs, tune test time estimates, and verify mechanical refactors with reproducible proofs.

Does Miles support low-precision and LoRA training?▼

Yes. It supports MXFP8, NVFP4, INT4, and BF16 training, plus LoRA and multi-LoRA adapters that load directly into SGLang for rollout.