SUBNET TRACKER · All subnets
SN120 · Affine
Every Bittensor Daily mention of SN120 Affine, newest first. Price changes are snapshots from each edition.
24h change +1.26%
- What happened
- SN120 (Affine) removed the z_bar gate relaxation, making the GREEN gate threshold flat at z > 1 going forward. Overnight, three candidate files (e526, e537, e540) passed the gate and moved the model head to a new checkpoint. The team outlined plans to raise the reward cap from 8k to 12k tokens, add an IRT term to the baseline draw, and support a step-multiplier hyperparameter for miners. A king gain multiplier went live from E581, scaling a new king's payout share by its measured pass-rate gain over its parent.
- Community
- A community member ran a holdout-based breakdown of pass/capped/wrong rates across checkpoints and shared the results with the team. Miners pressed the team on documentation inconsistencies between mining.json, GitHub, and an outdated llms.txt page, which the team addressed directly. Discussion also covered whether a step-multiplier or n_steps hyperparameter would speed training and whether switching to a per-submission queue fee would be fair to miners already waiting in line.
- Official
- The team corrected an earlier note, confirming the z_bar relaxation was removed and the gate is flat at z > 1. It announced plans to adjust the reward cap, add an IRT term, support a step multiplier, and fix the broken green publish, and separately detailed how the new king gain multiplier calculates and clamps payout shares.
24h change +0.57%
- What happened
- SN120 (Affine) replaced its mining client with a single-file upload format (affine-batch-v1), training one GRPO step per hotkey turn. The team acknowledged that the promotion gate's noise estimate had been too strict, effectively requiring about +3 points of improvement instead of the intended +1.5-2, and corrected this starting at epoch 354. Epoch 354 then went GREEN using a training file with a single rollout per prompt rather than multiple rollouts per prompt. The team later reported trainer improvements cutting epoch time from 80 to 24 minutes, a 3x reduction in variance, and head model accuracy reaching 90%.
- Community
- Miners questioned whether the epoch 354 win was based on legitimate GRPO training given the single-rollout-per-prompt format, and some asked whether the timing right after the gate fix was coincidental. Several participants described the competition as unfair or luck-based, with one calling for the subnet to be renamed "LOTTO" and others saying it had become a game for miners with well-timed hotkey submissions.
- Official
- The team explained the gate fix and published details of the epoch 354 winning file for review, confirming all 125 answers were correct against reference answers. It stated that single-rollout training is not standard GRPO but is a legitimate way to move the model, citing a held-out pass improvement from 0.643 to 0.728, and said it will publish the training script and post-epoch baseline numbers and probe results to let miners audit future rounds.
24h change −1.10%
- What happened
- SN120 (Affine) training epochs stalled for about 4 hours after the e135 training job exceeded its 2-hour time limit; the job was closed without a model update, keeping the checkpoint at 4e4157b2, and the trainer's time limit was doubled. The following e136 checkpoint reached a held-out task score of 0.875, up from 0.72 at e134 and 0.56 for the base model. Separately, Affine (SN120) launched Bittensor's first public AI video-generation studio along with a competition to optimize video generation efficiency.
- Community
- Several miners flagged the stalled epochs and asked when training would resume. Miners have repeatedly asked the team to share the private UID ban logic, saying high-scoring miners are receiving no emissions despite high FLIP scores. A community member offered a technical explanation for a possible scoring bug, pointing to the epoch optimizer code and suggesting epochs with insufficient data should be dropped or deferred.
- Official
- The team explained the stall was caused by the e135 job running past its time limit due to long rollouts, confirmed the checkpoint was not updated, and said the trainer's time limit was doubled before e136 began collecting normally. The team also confirmed the 0.875 held-out score was a genuine result rather than a display bug, while noting the small task set (32) could cause a few points of variance between epochs.
24h change −2.42%
- What happened
- SN120 (Affine)'s revised scoring mechanism concentrated emissions on two coldkeys across multiple epochs, with miners noting the same operators repeatedly took 10 of 10 paid slots. The team posted epoch-by-epoch updates for epochs 72 through 75, reporting held-out benchmark scores that moved between .594 and .75 while payouts remained fixed at 10 UIDs with nothing burned. Separately, the team relaunched Affine Studio, a video tool aimed at providing an ultra-fast video model, and introduced a new mech-1 competition track alongside the existing mech-0.
- Community
- Miners repeatedly asked the team to publish the scoring mechanism and per-UID scores, questioning how two coldkeys could hold nearly all top slots across consecutive epochs. Some also raised concerns about private commits in the new mech-1 track.
- Official
- The team posted epoch-by-epoch scoring updates explaining that some unchanged weight windows were due to scoring taking longer than usual, and said epoch 76 training would drop unresolved rollouts rather than train on them.
24h change +0.81%
- What happened
- SN120 (Affine)'s on-chain weight-setting froze overnight after the writer's scoring step began running past its time limit, causing stale weights and leaving the dashboard's network-data.json stuck on an earlier epoch; the team raised the limit and restored a fresh weight set. Separately, a community member reported that the submission logic was only accepting hotkeys with multiple leading zeros, which some miners were allegedly exploiting through repeated registrations. Training continued through this period, moving to 256 tasks per epoch from epoch 54 onward, with the held-out checkpoint score rising from .594 to .75 before dropping back to .656, which the team attributed to noise on a small 32-problem evaluation set.
- Community
- Several community members flagged the frozen weights and asked the team to check the dashboard, while a validator posted specifics on the owner validator's repeated resubmission of an identical weight vector and raised fairness concerns for newly registered miners who were receiving no weight. Another participant separately asked how miner weights are calculated if submissions are no longer verified, citing concerns about cheating.
- Official
- The team confirmed the freeze was caused by the scoring step exceeding its time limit, raised the limit, and reported a fresh weight set on chain with 244 UIDs including the newest registrations; a later update confirmed the dashboard was updating again through epoch 58.
24h change −1.78%
- 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.