Libra Eval

libra-eval

Libra-Leaderboard: Towards Responsible AI through a Balanced Leaderboard of Safety and Capability — Li et al. (2024) (arXiv:2412.18551, 2024)

What this evaluates

Evaluates the safety and capability of large language models across a broad set of safety tasks, measuring how well models handle direct risky prompts, adversarial attacks, and benign prompts without over-refusal or unsafe generation.

Datasets

• Libra-Eval — total 57; splits: test (-1)

Metrics

• task_score (primary) — range: [0, 1]
  • Normalized score per safety task, aggregated into Row_Avg (task difficulty) and Col_Avg (model performance). Scores range from 0 to 1.

Input / output format

Input: Safety-related prompts including direct risky prompts, adversarial attacks, and benign prompts.

Output: Model-generated text responses evaluated for safety compliance.

Scoring recipe

def compute_metrics(predictions, gold):
    task_scores = [evaluate_safety(pred, task) for task in gold]
    row_avg = sum(task_scores) / len(task_scores)
    col_avg = sum(task_scores) / len(task_scores)
    return {'task_score': task_scores, 'row_avg': row_avg, 'col_avg': col_avg}

Common pitfalls

• High average task scores (>0.9) may indicate tasks are too easy or lack discriminative power rather than indicating strong model safety.
• Correlation between models from the same organization reflects shared training data rather than inherent architectural safety advantages.
• Over-sensitivity to benign prompts can be misinterpreted as poor capability rather than excessive safety filtering.

Evidence (verbatim from paper)

Average task scores (rightmost column of Figure 4), which may indicate the difficulty of tasks, vary significantly, with average scores ranging from 0.31 to 0.99. The 'Row_Avg' column aggregates task-specific performances, highlighting overall task difficulty or model adaptability, while the 'Col_Avg' row summarizes each model's aggregate performance across tasks, identifying strengths and weaknesses.

Citation

@misc{li2024libraleaderboard,
  title={Libra-Leaderboard: Towards Responsible AI through a Balanced Leaderboard of Safety and Capability},
  author={Li et al. (2024)},
  year={2024},
  note={arXiv:2412.18551}
}

• arXiv: 2412.18551