Hosted RL

Train RL policies on cloud GPUs

Pick an environment — gymnasium classic control, Atari, MuJoCo, LLM RLHF, or robotics sim — a baseline repo, and a GPU. Metrics stream live. When the reward curve flattens, deploy the policy as a scale-to-zero cog or serverless endpoint.

Launch a training run OSS starter repos

CLI

app api POST /api/rl/jobs '{
  "envId": "cartpole-v1",
  "algoId": "ppo",
  "hardware": "gpu-rtx4090"
}'
app api GET  /api/rl/jobs/<job-id>          # poll metrics
app api POST /api/rl/jobs/<job-id>/deploy \
  '{"target": "serverless"}'                # policy endpoint

Per-second GPU billing from the shared credit balance; jobs scale to zero when training finishes.

1 · environment

Pick an environment

2 · algorithm repo

Pick a baseline

3 · gpu tier

Pick a GPU

Per-second billing with the platform margin included, the same catalog as Cog Studio and pricing.

live training metrics

Reward, loss, entropy, and throughput

Launch a run above to stream reward, loss, entropy, and throughput here, polled from /api/rl/jobs/{id}. Signed-out visitors get a simulated run.

open source

Starter repos

Every baseline here is an MIT-licensed repo. Each ships a single-file training loop, a cog.yaml, an app.nz policy server, and a serverless deploy manifest per docs/BUILDING_COGS.md.

MIT · PPO + DQN

appnz-rl-ppo-starter

Single-file PPO on gymnasium (classic control, Atari, MuJoCo) with cog.yaml, an app.nz policy server, and a serverless deploy manifest.

github.com/lee101/appnz-rl-ppo-starter
MIT · GRPO

appnz-rl-rlhf-lite

Minimal GRPO-style RLHF loop on a tiny LLM with programmable rewards, packaged for cog + serverless deploy on app.nz.

github.com/lee101/appnz-rl-rlhf-lite
MIT · TABULAR-Q + PPO

appnz-music-rl

Gymnasium-compatible piano composition environment, Q-learning trainer, R2 policy artifact, and Opus Cog inference.

github.com/lee101/app-site/tree/master/training/music

Any container that trains and then serves the cog HTTP contract can run here. Build it with hosted builds, register it on /cogs, and post metrics to the training API.

Train the policy, then serve it

Train on per-second billed GPUs, deploy the policy as a scale-to-zero endpoint, and call it over HTTP.

Launch a run Cog Studio Install the CLI