Configuration

BlockZero miners are configured via a YAML file. This page walks through every field.

Example Configuration

# Model
model_path: Qwen/Qwen3-VL-30B-Instruct  # or local path
expert_group: 0                           # your assigned expert group (0-indexed)

# Bittensor
chain_endpoint: wss://entrypoint-finney.opentensor.ai:443
netuid: 42                                # BlockZero's subnet UID
wallet:
  name: my-wallet
  hotkey: my-hotkey

# Storage
checkpoint_dir: ./checkpoints
data_path: /data/expert-group-0/train.jsonl

# Training
batch_size: 4                             # per-GPU batch size
gradient_accumulation_steps: 4           # effective batch = 4 × 4 = 16
learning_rate: 3e-5
warmup_steps: 100
max_steps: 500
fp16: true

# Monitoring
wandb: false
wandb_project: blockzero-miner

# Checkpoint management
checkpoint_keep_top_k: 3

# DataLoader
num_workers: 4
seed: 42

Field Reference

Model Settings

Field	Type	Default	Description
model_path	string	required	Path to base model or HuggingFace model ID
expert_group	integer	required	Your assigned expert group ID (0-indexed). Get this from the validator.

Bittensor Settings

Field	Type	Default	Description
chain_endpoint	string	wss://entrypoint-finney.opentensor.ai:443	Chain WebSocket endpoint
netuid	integer	required	BlockZero subnet UID
wallet.name	string	default	Bittensor coldkey name
wallet.hotkey	string	default	Bittensor hotkey name

Storage Settings

Field	Type	Default	Description
checkpoint_dir	string	./checkpoints	Where to save checkpoints (local directory).
data_path	string	required	Training dataset path (JSONL format, one example per line)

Training Settings

Field	Type	Default	Description
batch_size	integer	4	Per-GPU batch size. Reduce if you hit OOM.
gradient_accumulation_steps	integer	4	Steps to accumulate before optimizer step. Effective batch = batch_size × steps.
learning_rate	float	3e-5	AdamW learning rate. Start here; tune down if gradients are unstable.
warmup_steps	integer	100	LR warmup steps at the start of each training cycle
max_steps	integer	500	Max training steps per cycle before submission
fp16	boolean	true	Enable FP16 mixed precision. Nearly always better.

Monitoring

Field	Type	Default	Description
wandb	boolean	false	Enable Weights & Biases logging
wandb_project	string	blockzero-miner	W&B project name

Checkpoint Management

Field	Type	Default	Description
checkpoint_keep_top_k	integer	3	Number of checkpoints to retain (by lowest validation loss). Older ones are pruned.
num_workers	integer	4	DataLoader worker processes
seed	integer	42	Random seed for reproducibility

Multiple Expert Groups

If you are running multiple miners on the same machine (different expert groups), create a separate config file for each:

# Create separate config directories per expert group, then run each pair of processes:
# Group 0:
python mycelia/miner/train.py --path /path/to/checkpoints/miner/<hotkey>/group-0/
python mycelia/miner/model_io.py --path /path/to/checkpoints/miner/<hotkey>/group-0/

# Group 1:
python mycelia/miner/train.py --path /path/to/checkpoints/miner/<hotkey>/group-1/
python mycelia/miner/model_io.py --path /path/to/checkpoints/miner/<hotkey>/group-1/

Each miner should have a different checkpoint_dir to avoid conflicts.