diff --git a/ci/benchmarks/.benchmarks.yml b/ci/benchmarks/.benchmarks.yml index 6aaa892c..46b7760e 100644 --- a/ci/benchmarks/.benchmarks.yml +++ b/ci/benchmarks/.benchmarks.yml @@ -30,8 +30,8 @@ OPENAI_API_BASE: "https://api.openai.com/v1" artifacts: paths: - - cai/benchmarks/seceval/eval/outputs/ - - cai/benchmarks/cybermetric/outputs/ + - benchmarks/seceval/eval/outputs/ + - benchmarks/cybermetric/outputs/ expire_in: 12 month tags: - p40 @@ -53,8 +53,9 @@ benchmarks-test-cybermetric-openrouter: <<: *run_benchmarks - script: - - python3 benchmarks/cybermetric/CyberMetric_evaluator.py --model_name openrouter/qwen/qwen3-32b:free --file_path benchmarks/cybermetric/CyberMetric-2-v1.json + variables: + MODEL_NAME: "openrouter/qwen/qwen3-32b:free" + BENCHMARK_FILE: "benchmarks/cybermetric/CyberMetric-2-v1.json" benchmarks-test-cybermetric-openai: <<: *run_benchmarks