ConnitoAI

Training large language models typically requires centralized infrastructure, significant computational resources concentrated in one location, and long training timelines. Practitioners need ways to distribute training work across multiple machines and contributors while maintaining coordination and quality control, but existing distributed training frameworks lack economic incentives for reliable participation.

A decentralized, incentivized network of contributors can split LLM training across independent miners, each specializing in expert groups within a Mixture-of-Experts architecture. Validators coordinate the overall training process, aggregate contributions from all miners, and score each miner's work quality and efficiency. Economic rewards drive miners to contribute high-quality computations, creating a self-sustaining collaborative training environment without central authority.

Subnet 102 implements this model on Bittensor by distributing expert groups across miners using Hivemind for coordination. Validators run FastAPI servers to orchestrate training, aggregate miner outputs, expose performance metrics via Prometheus, and assign TAO rewards based on miner contribution scores. This enables practitioners to train large models collaboratively without needing monolithic hardware or centralized control.

Subnet 102 is a Bittensor subnet for collaborative, decentralized training of large language models using a Mixture-of-Experts (MoE) architecture. The network distributes model experts across independent miners, with validators coordinating the training process, aggregating contributions, and scoring miners to drive TAO rewards.