τeuτonic Subnet Discusses Model Training Challenges and Metrics

The τeuτonic team analyzed bottlenecks in model training convergence, validator evaluation speed, and dataset scaling issues. They reported failed trainability probes against the current 'king' model and shared benchmark metrics for MMLU, BBH, and other datasets.

• •Training speed limited by low convergence rate and large datasets requiring 8*B300 GPUs
• •Current 'king' model fails trainability probe with NaN loss errors for some users
• •Proposed optimizations include pre-tokenizing during training and threshold reduction
• •Shared model benchmarks show MMLU accuracy of 0.7668 and GSM8K exact match of 0.8347