Ralph

Clone the Ralph repository

git clone https://github.com/RalphLabsAI/ralph && cd ralph

Install core dependencies and chain SDK for on-chain operations

pip install -r requirements.txt -r requirements-chain.txt

Verify your compressed model meets the bit-budget requirement before committing

python -m eval.bitrate path/to/your/model.gguf

Dry-run the commit to see what will be written on-chain(optional)

python -m miner.submit commit --ckpt ./my-compressed-qwen3 --tier ternary --uri hf://<you>/<repo>@<commit-sha> --wallet <wallet> --hotkey <hotkey> --netuid 40 --dry-run

Seal the hash of your exact bytes on-chain before the round nonce exists

python -m miner.submit commit --ckpt ./my-compressed-qwen3 --tier ternary --uri hf://<you>/<repo>@<commit-sha> --wallet <wallet> --hotkey <hotkey> --netuid 40

Publish the artifact URI and salt after the round opens

python -m miner.submit reveal --ckpt ./my-compressed-qwen3 --wallet <wallet> --hotkey <hotkey> --netuid 40

Clone the Ralph repository

git clone https://github.com/RalphLabsAI/ralph && cd ralph

Install core dependencies and chain SDK

pip install -r requirements.txt -r requirements-chain.txt

Install GPU-specific dependencies for model scoring (L2/L3 audits)(optional)

pip install -r requirements-gpu.txt

Install system package for code execution sandbox

sudo apt-get install bubblewrap

Run a single audit pass at L0 (arithmetic verification, CPU-only)(optional)

python -m eval.auditor --once --require L0 --signer <validator-record-key>

Continuously audit new rounds at L0 and L1 (exam selection verification)(optional)

python -m eval.auditor --follow --require L0,L1 --signer <key> --interval 600

Audit with judge model inference (L2), requires GPU(optional)

python -m eval.auditor --follow --require L0,L1,L2 --signer <key> --observer <judge-hf-id>

Run as weight-setting validator: verify and set emission weights on-chain(optional)

python -m eval.auditor --follow --signer <key> --anchor-hotkey <scoring-validator-ss58> --wallet <yours> --hotkey <yours> --interval 1200 --set-weights