run.example.yaml

# Harbor run config for the WildAWM XAI task sets.
#
#   cp run.example.yaml run.yaml      # once, then set model_name
#   harbor run -c run.yaml            # no agent flags needed
#
# CLI flags override the file, e.g. switch the split with
#   harbor run -c run.yaml -p dataset/train-sft-100-set
# Inspect the resolved config with
#   harbor run -c run.yaml --print-config
jobs_dir: jobs
n_attempts: 4               # rollout attempts per task
n_concurrent_trials: 8      # concurrent trials
agent_setup_timeout_multiplier: 3   # npm install headroom inside the task image

agents:
  - name: opencode
    model_name: openai/<model>
    kwargs:
      version: 1.17.7       # opencode-ai CLI version, the one the benchmark was validated with

datasets:
  - path: dataset/val-33-set

environment:
  extra_docker_compose:
    - tools/host-gateway.yaml

extra_instruction_paths:
  - tools/contracts/completion_contract.md