run:
  name: exp1
  output_root: data/runs/eval_suite
runner: h200
runners:
  h200:
    tensor_parallel: 1
    gpu_memory_utilization: 0.92
    max_num_seqs: 128
    port: 8000
  two_h100:
    tensor_parallel: 2
    gpu_memory_utilization: 0.9
    max_num_seqs: 96
    port: 8000
  auto:
    tensor_parallel: auto
    gpu_memory_utilization: 0.92
    max_num_seqs: 128
    port: 8000
model:
  id: meta-llama/Llama-3.1-8B
  substrate: exp1
  max_model_len: 8192
  max_lora_rank: 128
  max_loras: auto
  reasoning_parser: none
  enable_thinking: false
  chat_template: auto
judge: anthropic/claude-sonnet-4-6
max_connections: 64
keep_serving: false
arms: []
suites:
  benign_agentic:
    type: inspect
    tasks:
    - name: benign_agentic
      task: why_gen/inspect_tasks/benign_agentic.py@benign_agentic
      epochs: 5
      temperature: 1.0
      max_tokens: 4096
      task_args:
        tool_format: am_xml