Alice Protocol

Decentralized AI training network for Alice-7B. Miners train local shards and earn ALICE in proportion to effective tokens.

Quick Start

git clone https://github.com/V-SK/Alice-Protocol.git
cd Alice-Protocol
./miner/bootstrap.sh --ps-url https://ps.aliceprotocol.org

The default miner mode is now Plan B.

Bootstrap will:

• detect your hardware
• install dependencies
• create or reuse ~/.alice/wallet.json
• download the current model if needed
• start the miner

To use your own wallet:

./miner/bootstrap.sh \
  --ps-url https://ps.aliceprotocol.org \
  --address YOUR_ADDRESS

Mining Model

Plan B is the default production path.

• The miner receives shard assignments from the network
• The runtime trains locally and uploads a compressed delta at epoch end
• Rewards are based on effective tokens, not a user-chosen batch size

Plan A is still available as a legacy fallback:

python3 miner/alice_miner.py --address YOUR_ADDRESS --mode plan_a

Hardware and Batch Assignment

Batch size is assigned by the network based on the miner's hardware profile. A manual --batch-size override can only reduce the assigned cap; it cannot increase it.

GPU VRAM	Assigned Batch Size	Tokens per Shard
16GB	1	10,240
24GB	4	40,960
32GB	8	81,920
48GB	16	163,840
80GB	32	327,680

Other requirements:

• RAM: 16GB+
• Disk: 50GB+
• Stable internet connection

Rewards

Reward accounting uses the assigned batch size:

effective_tokens = completed_shards * assigned_batch_size
tokens_trained = effective_tokens * 10240
your_share = your_effective_tokens / total_network_effective_tokens
your_reward = epoch_reward * 89% * your_share

This means larger GPUs earn proportionally more per shard because they train more tokens per shard.

Common Commands

Default Plan B:

python3 miner/alice_miner.py --address YOUR_ADDRESS

Separate Reward Address (Cloud GPU Safe Pattern)

Why:

• Use --address as the control/sign-in identity the miner registers with.
• Use --reward-address as the payout destination when the miner runs on rented or shared cloud GPUs.
• This keeps the reward wallet separate from the operational wallet you expose on the remote host.

How:

• If --reward-address is omitted, rewards go to --address.
• If --reward-address is set, rewards go to that address while the miner still authenticates as --address.
• The safe pattern is to keep --address on the cloud worker and direct rewards to a separate cold or payout wallet.

Example:

python3 miner/alice_miner.py \
  --address YOUR_CONTROL_ADDRESS \
  --reward-address YOUR_PAYOUT_ADDRESS

Multi-GPU:

CUDA_VISIBLE_DEVICES=0 ./miner/bootstrap.sh \
  --ps-url https://ps.aliceprotocol.org \
  --address YOUR_ADDRESS \
  --instance-id gpu0

CUDA_VISIBLE_DEVICES=1 ./miner/run_miner.sh \
  --ps-url https://ps.aliceprotocol.org \
  --address YOUR_ADDRESS \
  --instance-id gpu1

Background mode:

CUDA_VISIBLE_DEVICES=0 nohup ./miner/run_miner.sh \
  --ps-url https://ps.aliceprotocol.org \
  --address YOUR_ADDRESS \
  --instance-id gpu0 \
  > /tmp/miner_gpu0.log 2>&1 &

Links

• Website: https://aliceprotocol.org
• Explorer: https://aliceprotocol.org/explorer
• Wallet CLI: https://github.com/V-SK/alice-wallet

Notes

• Plan B is the default miner path.
• Plan A is deprecated and should only be used for compatibility or debugging.
• If a local model is too far behind, the miner will refresh the full model instead of staying on stale weights.