Self-play: 50 rounds, 0.1s sleep, max GPU utilization

Bumped from 20 rounds/tier to 50. Reduced sleep from 1s to 0.1s.
GPUs should run near 100% — Ollama queues requests internally.
mortdecai-sites container (CT 650) created on pve112.
Landing page live at mortdec.ai.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

training/scripts/overnight_selfplay.sh

@@ -8,7 +8,7 @@
 MODEL="${1:-mortdecai-v4}"
 RCON_HOST="${2:-192.168.0.244}"
 RCON_PORT="${3:-25578}"
-ROUNDS_PER_TIER=20
+ROUNDS_PER_TIER=50
 SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
 OUTPUT="$SCRIPT_DIR/../../data/processed/self_play.jsonl"
 

training/scripts/self_play.py

@@ -620,7 +620,7 @@ def main():
                 examples = trace_to_training(trace)
                 all_examples.extend(examples)
                 stats["training_examples"] += len(examples)
-                time.sleep(0.5)
+                time.sleep(0.1)
 
         elif tier == 2:
             # Self-critique: model generates prompt + response, RCON validates
@@ -654,7 +654,7 @@ def main():
                 examples = trace_to_training(trace)
                 all_examples.extend(examples)
                 stats["training_examples"] += len(examples)
-                time.sleep(1)
+                time.sleep(0.1)
 
         elif tier == 3:
             # Adversarial: generate prompts in Session A, respond in fresh Session B
@@ -703,7 +703,7 @@ def main():
                 examples = trace_to_training(trace)
                 all_examples.extend(examples)
                 stats["training_examples"] += len(examples)
-                time.sleep(1)
+                time.sleep(0.1)
 
         stats["rounds"] += 1