SUBNET TRACKER · All subnets
SN78 · Umi
Every Bittensor Daily mention of SN78 Umi, newest first. Price changes are snapshots from each edition.
24h change +8.00%
- What happened
- SN78 (Umi)'s owner resolved an extended debate over C6 scoring by deciding to use only the eight new evaluation clips rather than the original six, the combined 14, or a 50:50 split, with UID 172's model set as the baseline for C7. C6 scoring and certification were completed and corrected results published, though reward release to validators was still pending. C6 weights later went live on-chain, C7 submission intake opened with a planned 12-hour closing warning, and the subnet's ETAO-based emission split was changed to shift incentive from validators toward miners and the owner (owner 18→30.8%, validators 41→15.5%, miners 41→53.8%). The owner also said the per-bucket emission decay for C7 and later cohorts would move from 0.75 to 0.85, with a 1% grace margin added around the baseline.
- Community
- Miners debated whether the scoring change was fair to overfitted versus non-overfit models, with some calling the shifting rules confusing and others supporting the move to the new clips. A miner also proposed giving the top 3-5 models fixed, score-ordered reward shares to reduce large payout swings between closely scored models.
- Official
- The owner confirmed the final C6 scoring table would use the new 8 clips, announced the ETAO split and C6 weights going live, and outlined the C7 emission decay and grace-margin adjustments.
24h change +18.64%
- What happened
- SN78 (Umi) completed certification of all 103 Cohort 6 (C6) model submissions on Oct 26 and began reward activation on-chain. During review, the team found that several top-scoring C6 models had achieved high scores by memorizing public FLEURS-ASL and 2M-Flores-ASL test data rather than generalizing to new translation tasks. The team is now building a larger evaluation set with unseen data and re-running scoring for the top models before C6 weights are finalized; an initial pass already produced different rankings for several top-5 submissions. Separately, a miner reported a container CPU-limit bug causing MediaPipe segfaults that invalidated a model's continuous-translation score.
- Community
- Community members raised the scoring-formula inconsistency between C5 and C6 band weighting, and one member presented evidence that a specific UID's model had been trained on nearly all FLEURS-ASL and 2M-Flores-ASL data, scoring 0.90-0.997 on seen clips versus 0.15 on unseen ones. Members also questioned how to handle suspicious models and requested all submissions remain downloadable for review. The miner affected by the CPU-limit bug asked for their model to be included in the re-evaluation.
- Official
- The team confirmed the C6 scoring was "a bit broken" and is re-running evaluation with a larger, non-FLEURS dataset before setting C6 weights. The team also said upcoming changes are expected to increase miner emissions by roughly 31% of current levels, with a formal announcement to follow. The official SN78 account reported $22,471 traded on Base over two weeks, with 2.2% of circulating alpha bridged via ForeverMoney, 70% of it in the Aerodrome SN78/TAO pool, 348 trades since Oct 8, and $14.4K of liquidity added on Base.