Our mission is to maximize the collective training efforts of the Bittensor community by incentivizing the sharing of models, knowledge, insights, and code.

Main goals:

  1. Train and share state-of-the-art models and code to contribute to real-world applications.
  2. Incentivize both individual and collective effort.
  3. Maintain a consistent, high pace of high quality iterations.
  4. Uphold transparency and fairness among all stakeholders.

In more detail:

1. Train and share state-of-the-art models and code to contribute to real-world applications.

A major goal of coldint is to contribute something to the world. Miners are primarily incentivized for their contribution in the form of trained models that are publicly shared. A secondary incentive is given to miners or other contributors for sharing insights by contributing to the canonical miner codebase. This may range from pointing out bugs, to adding small features and optimizations, and to rewriting large parts of code.

2. Incentivize both individual and collective effort.

By rewarding small improvements, individual miners are incentivized to regularly share their improved work. This in turn enables other miners to build upon earlier contributions, accelerating the collective effort. Code contributions that enable combining individual training efforts into an even better combined model will be highly rewarded, as this will truly unlock the collective potential.

3. Maintain a consistent pace of high quality iterations.

Within coldint, iterations are done in three main areas:

  • models are improved continuously by miners
  • training objectives are scheduled to change periodically, at a steady pace
  • validator and miner codebases are continuously refined

All stakeholders can participate in improving each of these areas. This leads to a dynamic community which promotes the exchange of ideas, insights and knowledge.

4. Uphold transparency and fairness among all stakeholders.

Coldint strives to be transparent by publicizing miner code, training objective schedules, awarded bounties and research results. Coldint invites community members to suggest innovative training goals. Mitigating steps are taken to prevent well known “metagaming” tactics from earning any reward. If new gaming opportunities arise that would lead to misalignment between miner and subnet objective, bug bounties should ensure that reporting bugs is as lucrative as exploiting them. Bug bounties are rewards for code contributions, and will be publicized to elicit additional bug reports and fixes.

Further reading

Various subnet design choices made will be discussed in the Discord channel and in blog posts on this site.

Insights

Sharing our experiences, ideas, code, analysis, ...

Growing models

In this blog, and in our Discord channel, we discuss training in detail. A topic that is often overlooked, is how to grow a model. Especially in incentivized, collaborative and distributed training, this is a Read more…

On training (part 2)

This post discusses one of the essential ingredients for training a model: how to see whether your model is actually improving, in other words, how do you benchmark the result of your training? Please read Read more…

On incentive (part 2)

In the previous post on incentive we described how loss values are used to score models against each other, and how the relative standing between models changes over time due to decaying advantage. From that Read more…