README.md

Samantha Thought-Loop SFT

• Base model: google/t5gemma-l-l-prefixlm-it
• Cleaned dataset repo: BRlkl/timed_natural_conversations_5min4
• Raw dataset repo: BRlkl/timed_natural_conversations_5min4
• Latent slots: 256
• Fixed tick seconds: 1.0
• Chunk ticks: 300
• Max horizon ticks: 300
• Explicit time features: False
• Gate loss mode: hazard
• Hazard soft target window: 16
• Hazard soft target tau ticks: 4.0
• Hazard censor weight: 0.25
• TBPTT by bucket: {}

Best Validation Metrics

• active_ticks: 1888.000000
• decoder_token_count: 2494.000000
• gate_accuracy: 0.945445
• gate_f1: 0.000000
• gate_positive_weight: 0.000000
• gate_precision: 0.000000
• gate_pred_positive_count: 0.000000
• gate_recall: 0.000000
• gate_target_positive_count: 103.000000
• hazard_censor_count: 12.000000
• hazard_event_count: 103.000000
• loss_decoder: 3.760284
• loss_gate: 3.530689
• loss_gate_ranking: 0.000000
• loss_state_delta: 0.257930
• loss_total: 6.434094