def run_eval()

in genai-on-vertex-ai/gemini/model_upgrades/instruction_following/vertex_script/eval.py [0:0]

• 20 lines of code
• 2 McCabe index (conditional complexity)

def run_eval(experiment_name: str, baseline_model: str, candidate_model: str, prompt_template_local_path: str, dataset_local_path: str, metric_name: str):
    timestamp = f"{datetime.now().strftime('%b-%d-%H-%M-%S')}".lower()
    prompt_template = open(prompt_template_local_path).read()
    
    metrics = EvalTask(
      dataset=load_dataset(dataset_local_path),
      metrics= [
        PairwiseMetric(
            metric=metric_name,
            metric_prompt_template=MetricPromptTemplateExamples.Pairwise.INSTRUCTION_FOLLOWING.metric_prompt_template,
            # Baseline model for pairwise comparison
            baseline_model=GenerativeModel(baseline_model),
        ),
    ],
      experiment=experiment_name
  ).evaluate(
      model=GenerativeModel(candidate_model),
      prompt_template=prompt_template,
      experiment_run_name=f"{timestamp}-{candidate_model.replace('.', '-')}"
  )
    
    print("Baseline model win rate:", round(metrics.summary_metrics[f'{metric_name}/baseline_model_win_rate'],3))
    print("Candidate model win rate:", round(metrics.summary_metrics[f'{metric_name}/candidate_model_win_rate'],3))