Train customer jobs
When a customer trains on the open tier, network miners train the model. A customer-job miner runs a queue worker: it claims a job, trains a z3-mini adapter on the job's training split, and submits it. Zils evaluates the result on data the miner never sees and serves accepted models through its API.
This is separate from the soup pools. zils-miner commits soup bodies; it has no customer-job command. One host can do both, and customer jobs then take priority on a shared card.
Who can join
Registering a hotkey on mainnet subnet 127 is the admission. There is no application and no operator approval. The coordinator at https://training.zils.ai reads the subnet's registrations itself. GET https://training.zils.ai/v1/config shows admission: mode chain, the network and netuid, and the terms_sha256 of the current miner terms.
You need:
- A hotkey registered on subnet 127 (Register on Bittensor). Keep the coldkey off the mining host.
- Linux x86_64 or WSL 2 with an NVIDIA GPU that supports BF16 and its CUDA driver. The worker needs 12 GB (12,288 MiB) of free GPU memory; training peaks at about 10.5 GB on an RTX 4090. CPU and Apple MPS are not supported.
- About 25 GB of free disk for models and dependencies, plus room for job data.
Customer jobs run on z3-mini only (d3-mini-4.5b-v1, inputs up to 8,192 tokens).
What you receive
For each job: the training split only (states, questions and labels) and, for an image job, its training photos. Calibration and test splits, acceptance thresholds and the customer's identity stay with Zils. Private-tier jobs never reach network miners.
The miner terms apply. Use the data only for the job, do not serve or hand over the adapter, and delete what you keep.
The worker deletes a job's training rows, photos and partial downloads as soon as it leaves the job: after it submits, when the lease is lost, when the job is cancelled or fails, and when it hands the job back. Anything a crash left behind is deleted before the next claim. The trained candidate, raw training output and logs stay on disk. The terms require you to delete those within 7 days of the job finishing; the worker does not do that yet.
1. Install
The miner installer clones the zils repository at the stable release into ~/.zils/zils and builds both runtimes. ZILS_SKIP_SETUP=1 skips the soup miner's setup questions:
curl -fsSL https://zils.ai/install.sh | ZILS_SKIP_SETUP=1 sh
cd ~/.zils/zils
export ZILS_D3_RUNTIME_PYTHON="$PWD/.venv-d3/bin/python"
.venv-d3/bin/python -m zils.d3 reference --out models/d3-referenceThe reference command downloads and checks the pinned z3-mini weights (about 9 GB). Run every command below from ~/.zils/zils.
2. Accept the terms and qualify the host
Read the miner terms. Mining with a registered hotkey is acceptance. Check that the hash of your copy matches the one the coordinator publishes:
sha256sum docs/mining-terms.md
curl -s https://training.zils.ai/v1/configThen qualify the host on its real hardware. This runs the pinned runtime check, a training step at the 8,192-token cap and an adapter reload, and writes qualification.json:
.venv-kev/bin/python -m scripts.qualify_node --model d3-mini-4.5b-v1 \
--runtime-python .venv-d3/bin/python --out .private/qualification \
--terms-sha256 YOUR_TERMS_SHA2563. Configure the worker
mkdir -p .private && chmod 700 .privateSave this as .private/queue-miner.json, with your own wallet details and the terms hash:
{
"coordinator": "https://training.zils.ai",
"hotkey": "YOUR_PUBLIC_HOTKEY_SS58_ADDRESS",
"wallet": {
"name": "YOUR_WALLET_NAME",
"hotkey": "YOUR_HOTKEY_NAME",
"path": "/absolute/path/to/your/wallets"
},
"terms_sha256": "YOUR_TERMS_SHA256",
"qualification": ".private/qualification/qualification.json",
"graded_scheduling": true,
"graded_min_free_mib": 12288
}chmod 600 .private/queue-miner.jsonThe wallet directory must hold the named hotkey on this machine; the worker checks that its address matches hotkey before it starts. graded_scheduling is required: it makes the worker send a signed heartbeat every 15 seconds, which is how the coordinator sees that it is online and has free GPU memory. Without it the worker is never offered a job, qualification runs included. Use a dedicated operating-system account for the miner.
4. Start the worker
export ZILS_D3_RUNTIME_PYTHON="$PWD/.venv-d3/bin/python"
.venv-kev/bin/python -m miner.queue \
--config .private/queue-miner.json --state .private/queue-miner \
--reference models/d3-reference --device cudaAt startup the worker admits itself: it sends the terms hash and the qualification report, signed by the hotkey. Every claim then carries the terms hash and the installed release. The worker checks for work every ten seconds and can stay quiet for a long time when there is none. Keep the same --state folder across restarts; only one worker can use it at a time. Use a service supervisor for continuous operation.
Work arrives in this order:
- Qualification jobs. Synthetic jobs that grade the host. Zils pays for up to three qualification runs per hotkey every 30 days, queued one at a time and only while your worker is online and sending heartbeats. They earn nothing.
- Customer jobs, once three qualification reports pass.
A successful submission prints Submitted candidate for job .... That means the artifacts arrived; Zils' evaluation decides whether the model is accepted.
Stay on the current release
Every claim names the zils commit the worker runs. Customer jobs go only to accepted releases: the current stable release (the commit in release/stable) and, for 24 hours after a required release, the one before it. After that, claims from the older release are refused, and the worker logs:
This miner is at ...; customer jobs need a current release. Run zils-miner update.Run zils-miner update, then restart the worker. An optional release refuses nothing; zils-miner doctor warns until you update. Each release says which it is in its notes on the releases page. A worker started from a copy without git history sends no release and gets no customer jobs.
How customer jobs pay
The adapter track has 20% of emissions (pool shares). A validator that runs the published rule pays it from signed platform evidence of paid API use:
- You earn when an open-tier model you trained is accepted, delivered to the customer, and the customer pays to use it through the API. That spend counts for 30 days after delivery and is settled over 72-hour windows.
- Miners as a group earn at most half of that spend (
revenue_cap0.5). The rest of the share burns. - Free and test use, qualification runs and undelivered models earn nothing. Only hotkeys registered on the subnet earn.
The full rule is in Run a validator. No earnings are promised: there may be no paid use, and your costs are yours.