Vicuna Benchmark Eval

---
name: vicuna-benchmark-eval
description: This benchmark assesses general language model capabilities and safety across diverse tasks like Fermi problems, roleplay, and coding. It evaluates how well models balance helpfulness, accuracy, and safety on non-safety-specific queries. Use when the user wants to benchmark on Vicuna_Benchmark, or asks about evaluating this task. Reports Net Win Rate.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2403.11838
  bibtex_key: luo2024guidealign
  confidence: high
---

# vicuna-benchmark-eval

> Ensuring Safe and High-Quality Outputs: A Guideline Library Approach for Language Models — Luo et al. (2024) (arXiv:2403.11838, 2024)

## What this evaluates

This benchmark assesses general language model capabilities and safety across diverse tasks like Fermi problems, roleplay, and coding. It evaluates how well models balance helpfulness, accuracy, and safety on non-safety-specific queries.

## Datasets

- **Vicuna_Benchmark** — total 80; splits: test (80)

## Metrics

- `Net Win Rate` **(primary)** — range: [-1, 1]
  - Calculated as (Win - Lose) / (Win + Lose) based on pairwise comparisons judged by GPT-4 across helpfulness, relevance, accuracy, detail, and safety.

## Input / output format

**Input**: 80 diverse queries covering various domains and task types.

**Output**: Text responses generated by the LLM.

## Scoring recipe

```python
pairs = generate_pairwise_prompts(model_a, model_b)
judgments = gpt4_judge.evaluate(pairs, criteria=['helpfulness', 'relevance', 'accuracy', 'detail', 'safety'])
wins = count(judgments == 'A wins')
loses = count(judgments == 'B wins')
net_win_rate = (wins - loses) / (wins + loses)
```

## Common pitfalls

- GPT-4 judge may favor verbose or highly detailed responses over concise ones.
- Safety is added as an extra dimension to a primarily capability-focused benchmark, which can skew results if models prioritize safety over helpfulness.

## Evidence (verbatim from paper)

> Employing this benchmark, we assess the efficacy of our approach across a spectrum of tasks extending beyond security-related scenarios. Following the methodology articulated by Chiang et al. (2023), we employ GPT-4 to evaluate LLMs’ responses based on criteria such as helpfulness, relevance, accuracy, and level of detail. Moreover, we incorporate safety as an additional evaluative dimension. In each experimental set, "Win", "Tie" and "Lose" refer to the outcomes on the left relative to the right of the "vs." in the corresponding caption. The value of "Net Win Rate" is calculated by dividing the difference between the value of "Win" and "Lose" for each row by the sum of the values for that row.

## Citation

```bibtex
@misc{luo2024guidealign,
  title={Ensuring Safe and High-Quality Outputs: A Guideline Library Approach for Language Models},
  author={Luo et al. (2024)},
  year={2024},
  note={arXiv:2403.11838}
}
```

- arXiv: 2403.11838