run:
  name: qwen35-9b-exp1-exp2-union
  output_root: data/runs/eval_suite
runner: h100
runners:
  h100:
    tensor_parallel: 1
    gpu_memory_utilization: 0.9
    max_num_seqs: 64
    port: 8000
  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: Qwen/Qwen3.5-9B
  path: null
  substrate: qwen35_9b
  max_model_len: 32768
  max_lora_rank: 128
  max_loras: auto
  reasoning_parser: qwen3
  enable_thinking: true
  chat_template: auto
judge: anthropic/claude-sonnet-4-6
max_connections: 64
keep_serving: false
arms: []
suites:
  preference:
    type: inspect
    tasks:
    - name: released_judge
      task: why_gen/inspect_tasks/preference.py@preference
      temperature: 0.0
      max_tokens: 2048
      task_args:
        kind: released
    - name: released_letter2
      task: why_gen/inspect_tasks/preference.py@preference
      temperature: 0.0
      max_tokens: 1024
      task_args:
        kind: released-letter2
  idqa:
    type: inspect
    tasks:
    - name: spec_open_qa
      task: why_gen/inspect_tasks/idqa.py@idqa
      temperature: 0.0
      max_tokens: 4096
  capability:
    type: inspect
    tasks:
    - name: arc_challenge
      task: inspect_evals/arc_challenge
      limit: 200
    - name: truthfulqa
      task: inspect_evals/truthfulqa
      limit: 200
    - name: gsm8k
      task: inspect_evals/gsm8k
      limit: 200
    - name: ifeval
      task: inspect_evals/ifeval
      limit: 200
  agentic:
    type: inspect
    cwd: /workspace/mats_project/code/external/model_spec_midtraining
    pythonpath: /workspace/mats_project/code/external/model_spec_midtraining
    tasks:
    - name: murder_explicit_america
      task: evals/agentic_misalignment
      epochs: 5
      temperature: 1.0
      max_tokens: 8192
      model_args:
        responses_api: false
      task_args:
        scenario: murder
        goal_type: explicit
        goal_value: america
        urgency_type: replacement
        prod: true
  benign_agentic:
    type: inspect
    tasks:
    - name: am_xml
      task: why_gen/inspect_tasks/benign_agentic.py@benign_agentic
      epochs: 5
      temperature: 1.0
      max_tokens: 4096
      task_args:
        tool_format: am_xml
    - name: json
      task: why_gen/inspect_tasks/benign_agentic.py@benign_agentic
      epochs: 5
      temperature: 1.0
      max_tokens: 4096
      task_args:
        tool_format: json