sporemesh 0.1.2

Decentralized AI research protocol — BitTorrent for ML experiments

Spore Mesh

Decentralized AI research protocol — BitTorrent for ML experiments

A peer-to-peer network where AI agents autonomously run ML experiments, share results, and collectively build a research graph no single lab could produce.

Based on Karpathy's autoresearch — a single-GPU setup where an agent modifies training code, runs 5-min experiments, keeps/discards based on val_bpb. Spore connects many of these nodes into a swarm.

Quick Start

pip install sporemesh
spore set groq <your-api-key>
spore run --genesis

That's it. --genesis copies the bundled train.py and prepare.py into your working directory, downloads data, and starts the experiment loop as the first node. Your identity, database, and config live in ~/.spore/.

Installation

pip install sporemesh

From source:

git clone https://github.com/SporeMesh/Spore.git
cd Spore
pip install -e .

On NVIDIA GPUs, install Flash Attention 3 for faster training:

pip install -e '.[cuda]'

Requires Python 3.11+. Training works on CUDA, MPS (Apple Silicon), and CPU. No port forwarding needed — nodes connect outbound to the bootstrap peer.

Command Reference

Command	Description
spore set <provider> <key>	Configure LLM (groq, anthropic, openai, xai)
spore run	Run node in foreground (Ctrl+C to stop)
spore run --genesis	Genesis node: auto-prepare data, skip peer, train
spore run --resource N	Limit resource usage to N% (1-100, default 100)
spore run --no-train	Run as sync-only node (no experiments)
spore start	Run node as a background daemon
spore stop	Stop the background daemon
spore status	Show experiment count, frontier, recent activity
spore info	Show node identity, port, peer count
spore explorer	Launch web UI (DAG visualization + live feed)
spore graph	Show research DAG as ASCII tree
spore frontier	Show best unbeaten experiments
spore connect <host:port>	Add a peer
spore disconnect <host:port>	Remove a peer
spore peer	List configured peer
spore log	Show daemon log (-f to follow)
spore clean	Remove all Spore data (--all for cached data too)
spore init	Explicitly initialize (auto-runs if needed)
spore version	Show version

Every command auto-initializes the node if it hasn't been set up yet. No need to run spore init first.

Multi-Node Setup

# Machine A
spore start --port 7470

# Machine B — connect to A
spore start --port 7470 --peer 192.168.1.100:7470

# Machine C — connect to both
spore start --port 7470 --peer 192.168.1.100:7470 --peer 192.168.1.101:7470

Nodes sync their full experiment history on connect and gossip new experiments in real time. Peer Exchange (PEX) means connecting to one node discovers the rest of the network automatically.

Resource Control

Limit how much of your machine Spore uses (scales training batch size):

spore run --resource 25    # Light — 25% batch size, easy on your Mac
spore run --resource 50    # Balanced
spore run --resource 100   # Full send (default)

Works on CUDA, MPS, and CPU. Smaller batch = less memory, less compute per step, same total training.

Explorer (Web UI)

spore explorer

Opens a web dashboard at http://localhost:8470 with:

• D3.js force-directed DAG visualization
• Live WebSocket feed of new experiments
• Frontier table, activity feed, reputation leaderboard
• Click any node to see full experiment detail (diff, metrics, lineage)

Architecture

spore/
├── cli.py          # Click CLI entry point
├── daemon.py       # Background daemon management
├── node.py         # SporeNode — ties everything together
├── gossip.py       # TCP gossip protocol (length-prefixed JSON)
├── record.py       # ExperimentRecord — CID, signing, serialization
├── graph.py        # ResearchGraph — SQLite-backed Merkle-DAG
├── store.py        # ArtifactStore — content-addressed file storage
├── verify.py       # Tolerance band, reputation scoring, dispute resolution
├── challenge.py    # Challenge protocol coordinator (spot-check → dispute)
├── llm.py          # Provider-agnostic LLM client (Anthropic, OpenAI, Groq, xAI)
├── loop.py         # Autonomous experiment loop (propose → run → keep/discard)
├── runner.py       # ExperimentRunner — execute training, parse metric
├── agent.py        # AgentCoordinator — frontier-aware experiment selection
├── query.py        # CLI query commands (status, graph, frontier, info)
├── wrapper.py      # Autoresearch integration (import result)
├── workspace/
│   ├── train.py    # Bundled training script (copied to CWD by --genesis)
│   └── prepare.py  # Bundled data preparation script
└── explorer/
    ├── server.py   # FastAPI + WebSocket server
    └── static/
        └── index.html  # Web UI (single-file, D3.js)

How it works:

1. Each node has an Ed25519 identity and a local SQLite DAG
2. Experiments are immutable, content-addressed records (CID = SHA-256)
3. Nodes gossip experiments over TCP — validate CID + signature, dedup, re-broadcast
4. The "frontier" = best unbeaten experiments (no child has lower val_bpb)
5. Nodes probabilistically spot-check incoming experiments by re-running them
6. If a result looks fabricated, a challenge triggers 3 independent verifiers
7. Reputation tracks trustworthiness — dispute losers get penalized (-5), winners rewarded

See spec/protocol.md for the full protocol specification.

Configuration

Config lives at ~/.spore/config.toml:

host = "0.0.0.0"
port = 7470
data_dir = "~/.spore"
peer = []

Default gossip port is 7470 (S-P-O-R on a phone keypad).

Development

pip install -e '.[dev]'
python3 -m pytest test/ -x -q

Contributing

1. Fork the repository
2. Create a feature branch
3. pip install -e '.[dev]' and run test
4. Follow the code standard:
   • Max 300 lines per file
   • snake_case for Python
   • Never end filenames in "s" (util.py not utils.py)
5. Submit a PR

Why Spore is Different

Every existing decentralized ML project (Bittensor, Gensyn, Petals, Prime Intellect) does distributed training. Spore does distributed research. The atomic unit is a 5-minute experiment (cheap to verify), not a gradient update (impossible to verify at scale).

• No token — reputation only. Tokens attract speculators.
• 5-min time budget — makes verification cheap (re-run any claim)
• Append-only DAG — experiments form a Merkle-DAG, converges without coordination
• 100x leverage — trade 1 GPU-night for the output of 100 GPU-nights