Affine
- What happened
- Affine (SN120) published migration code for a new MiMo trainer and later cut over to the MiMo-V2.6-Distill-Qwen-9B checkpoint with a new GRPO objective at global epoch 189, requiring miners to upgrade their runtime before continuing to mine. The team reported the first public MiMo epochs, with the signed fixed32 evaluation moving from 18/32 to 19/32 and then to 21/32 across successive updates, while cautioning this reflects answer-formatting corrections rather than confirmed learning gains. An audit flagged some closed MiMo task groups that passed grading checks but disagreed with the prescribed cached sampler at a token boundary, prompting an ongoing investigation. Separately, SN120 lowered min_difficulty and cut max_validators from 64 to 1, and enabled proof-of-work registration allowing miners to register with zero TAO.
- Community
- Miners and the team discussed why GRPO training had stalled, pointing to a small math dataset causing repeated rollout submissions, registration costs that favor miners with more slots, and concerns that miners prioritize TAO rewards over actual model improvement. Suggestions included expanding the dataset and requiring miners to submit batches from their first N trajectories rather than cherry-picked samples.
- Official
- The team published the MiMo migration code, confirmed the cutover at epoch 189, reported audit diagnostics on the cached-sampler mismatch, and noted that score corrections for answer formatting do not change the official signed 18/32 to 19/32 results.
In debate · Affine training stability and reward mechanism · 11 participants
-
The dataset is too small, causing repeated rollout submissions across epochs that stall training.
“dataset is too small, so we keep seeing the same submissions repeated every epoch.”
-
High registration costs favor miners with more slots and create a barrier for new miners to join.
“Right now, it's hard for new miners to join the competition because miners with more slots create a high barrier to reg war.”
-
Miners are focused on earning TAO rather than contributing to real model improvement.
“current miners care more about earning TAO than improving the model.”