# Brace Main Eval

> Evaluates the ability of audio-language models to align audio with captions and distinguish caption quality across different generation sources (human-human, human-machine, machine-machine). It probes fine-grained semantic and syntactic alignment capabilities under realistic captioning conditions. Use when the user wants to benchmark on BRACE-Main, or asks about evaluating this task. Reports F1-score.

- Skill: `qhjqhj00/brace-main-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/brace-main-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/brace-main-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Productivity
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/brace-main-eval

---


# brace-main-eval

> BRACE: A Benchmark for Robust Audio Caption Quality Evaluation — Guo et al. (2025) (arXiv:2512.10403, 2025)

## What this evaluates

Evaluates the ability of audio-language models to align audio with captions and distinguish caption quality across different generation sources (human-human, human-machine, machine-machine). It probes fine-grained semantic and syntactic alignment capabilities under realistic captioning conditions.

## Datasets

- **BRACE-Main** — total ?; splits: test (-1); repo https://github.com/HychTus/BRACE_Evaluation

## Metrics

- `F1-score` **(primary)** — range: [0, 1]
  - Harmonic mean of precision and recall for caption quality alignment or preference prediction between caption pairs.

## Input / output format

**Input**: Audio clip paired with two captions (Human-Human, Human-Machine, or Machine-Machine pairs) for quality comparison or alignment scoring.

**Output**: Model outputs a similarity score, preference choice, or quality rating for the caption pair.

## Scoring recipe

```python
def compute_f1(predictions, gold):
    tp = sum(1 for p, g in zip(predictions, gold) if p == 1 and g == 1)
    fp = sum(1 for p, g in zip(predictions, gold) if p == 1 and g == 0)
    fn = sum(1 for p, g in zip(predictions, gold) if p == 0 and g == 1)
    precision = tp / (tp + fp) if (tp + fp) > 0 else 0
    recall = tp / (tp + fn) if (tp + fn) > 0 else 0
    return 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0
```

## Common pitfalls

- CLAP models struggle with fine-grained Human-Human and Machine-Machine caption comparisons, often failing to identify subtle quality differences.
- Position bias in LALMs causes them to favor caption_0 or caption_1 regardless of actual content, especially under complex prompts.

## Evidence (verbatim from paper)

> On BRACE-Main, the best-performing model LAION-CLAP achieves an F1-score of 70.01, while others range from ~55 to 70 depending on architecture and subset.

## Citation

```bibtex
@misc{guo2025brace,
  title={BRACE: A Benchmark for Robust Audio Caption Quality Evaluation},
  author={Guo et al. (2025)},
  year={2025},
  note={arXiv:2512.10403}
}
```

- arXiv: 2512.10403