Votes and training

Each round, anyone can propose a training job: a title, a base model and a mix of datasets from the index. Wallets sign one vote per round. When the round closes, the proposal with the most weight becomes the next job in the queue.

Round 1

Open from 26 October 2026, 00:00 UTC to 9 November 2026, 00:00 UTC. Base models on the board: Qwen2.5 0.5B and 1.5B Instruct, Llama 3.2 1B Instruct, SmolLM2 1.7B Instruct, all pulled from Hugging Face.

Vote weight

Votes are signed messages, so voting costs no gas. Weight is read at tally time: one per wallet until staking opens; then the staked balance of $NODAL in the staking contract. Staked $NODAL also receives the 10% stake share of every paid query, paid in USDG.

The job

The winning mix is exported as prompt and answer pairs built from its chunks. A training node fine-tunes a LoRA adapter (rank 16, alpha 32, learning rate 2e-4, loss masked to the answer) and uploads it. The adapter records the datasets and their chunk counts.

Model usage pays the data

When a trained adapter is served, each paid call is split like a query: 70% to the owners of the datasets in its mix, pro rata to the chunks each one contributed, 20% to the serving node, 10% to the stake pool.