docker-compose-eval.yml

# Evaluation Runner (connects to shared verifiers)
#
# This file defines the sec-code-bench evaluation service that connects to
# the shared verifier network. Multiple instances can run in parallel.
#
# Prerequisites:
#   - Verifiers must be running: docker compose -f docker-compose-verifiers.yml up -d
#   - The 'secbench-verifier-net' network must exist
#
# Usage:
#   # Run multiple evaluations in parallel with different project names
#   docker compose -f docker-compose-eval.yml -p eval-model-a up
#   docker compose -f docker-compose-eval.yml -p eval-model-b up
#
#   # Or with different config files
#   CONFIG_FILE=config-model-a.yaml docker compose -f docker-compose-eval.yml -p model-a up
#   CONFIG_FILE=config-model-b.yaml docker compose -f docker-compose-eval.yml -p model-b up
#
# Environment variables:
#   - CONFIG_FILE: Path to config.yaml (default: ./config.yaml)
#   - LOCAL_RESULT_DIR: Directory for results (default: ./results)

services:
  sec-code-bench:
    image: sec-code-bench:v2.2
    build:
      context: .
      dockerfile: Dockerfile
    pull_policy: never
    entrypoint: ["/entrypoint.sh"]
    volumes:
      - ./docker/entrypoint.sh:/entrypoint.sh:ro
      - ${CONFIG_FILE:-./config.yaml}:/sec-code-bench/config.yaml:ro
      - ${LOCAL_RESULT_DIR:-./results}:/dockershare
    networks:
      - secbench-verifier-net
    restart: no

networks:
  secbench-verifier-net:
    external: true
    name: secbench-verifier-net