With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
Refinery is Bittensor Subnet 125, an incentivised research network for improving the algorithms behind language-model training. Its goal is to help open models learn more from the compute available to them.
Miners submit optimizer code. Validators train models with it and independently score the resulting checkpoints under a fixed compute budget. Improvements that pass confirmation earn rewards.
Why optimizers first
An optimizer decides how a model learns from each batch of data. A better update rule can reach a lower validation loss with the same compute, or reach a useful loss sooner. That makes it a concrete, measurable first target: contributors change one component and measure its effect under a shared training contract.
Rewards follow progress
Refinery is Bittensor Subnet 125, an incentivised research network for improving the algorithms behind language-model training. Its goal is to help open models learn more from the compute available to them.
Miners submit optimizer code. Validators train models with it and independently score the resulting checkpoints under a fixed compute budget. Improvements that pass confirmation earn rewards.
Why optimizers first
An optimizer decides how a model learns from each batch of data. A better update rule can reach a lower validation loss with the same compute, or reach a useful loss sooner. That makes it a concrete, measurable first target: contributors change one component and measure its effect under a shared training contract.
Rewards follow progress
What exists today
The open-source Refinery repository includes a mining guide, an optimizer template, instructions for LLM research agents, and the validator training and scoring pipeline. Training runs are capped at 188,000 steps or the compute budget, whichever comes first, and the submitted optimizer controls its own learning-rate schedule.
Who it is for
Optimizer specialists, ML engineers and research-agent teams, giving them a direct route into Bittensor mining through useful, reproducible discoveries.
What exists today
The open-source Refinery repository includes a mining guide, an optimizer template, instructions for LLM research agents, and the validator training and scoring pipeline. Training runs are capped at 188,000 steps or the compute budget, whichever comes first, and the submitted optimizer controls its own learning-rate schedule.
Who it is for
Optimizer specialists, ML engineers and research-agent teams, giving them a direct route into Bittensor mining through useful, reproducible discoveries.
Team
Refinery is developed by the Refinery team. Public development on the repository is by the GitHub account Barbariandev.
Subnet ownership
Subnet 125 now carries the Refinery identity under its current owner. It is a different project from the one previously featured on this page.
Community
No official X account or Discord server is listed, so community discussion takes place in the main Bittensor Discord.
Team
Refinery is developed by the Refinery team. Public development on the repository is by the GitHub account Barbariandev.
Subnet ownership
Subnet 125 now carries the Refinery identity under its current owner. It is a different project from the one previously featured on this page.
Community
No official X account or Discord server is listed, so community discussion takes place in the main Bittensor Discord.