burn_dragon 🔥🐉

burn_dragon is the dragon model + training workspace.

it pairs the dragon model stack with burn_p2p for native + browser p2p training, deployment, and live network operation.

the model shape follows the dragon hatchling / bdh paper.

what is here

• crates/burn_dragon_core: core model, state, and config
• crates/burn_dragon_language: language training + inference adapters
• crates/burn_dragon_p2p: p2p runtime, browser ui, deployment, and integration tests
• xtask: build, smoke, deploy, and release helpers

common paths

• model + language code: crates/burn_dragon_core, crates/burn_dragon_language
• p2p + deployment: crates/burn_dragon_p2p, crates/burn_dragon_p2p/deploy/README.md
• protocol/runtime layer: burn_p2p

quick start

cargo run -p xtask -- local-browser-e2e
cargo run -p xtask -- smoke
cargo run -p xtask -- deploy-check

Use local-browser-e2e as the first browser/p2p production-parity gate. It runs the deployment config drift checks, a local edge/auth/browser training receipt e2e, and the smallest real Chrome/WebGPU browser training smoke without forcing the full CI build matrix.