[Unit]
Description=Elizabeth Training Service %I
After=network.target
Requires=elizabeth-memory.service

[Service]
Type=simple
User=root
WorkingDirectory=/data/adaptai
Environment=PYTHONPATH=/data/adaptai/aiml/datascience
ExecStart=/usr/bin/python3 /data/adaptai/aiml/datascience/fast_training_pipeline.py \
  --model_name_or_path /workspace/models/qwen3-8b \
  --output_dir /data/adaptai/checkpoints/elizabeth-${%I} \
  --dataset_dir /data/adaptai/corpus-data/elizabeth-corpus/ \
  --num_train_epochs 1 \
  --per_device_train_batch_size 4 \
  --gradient_accumulation_steps 16 \
  --learning_rate 1.0e-5 \
  --max_seq_length 4096 \
  --save_steps 500 \
  --logging_steps 10 \
  --bf16 \
  --gradient_checkpointing

Restart=on-failure
RestartSec=30
TimeoutStopSec=300

# Memory and resource limits
MemoryMax=120G
CPUQuota=400%
IOWeight=100

# Security
NoNewPrivileges=yes
ProtectSystem=strict
ProtectHome=yes
PrivateTmp=yes

[Install]
WantedBy=multi-user.target