Jamendo Mt Qa Eval

---
name: jamendo-mt-qa-eval
description: Evaluates audio-language models on multi-track comparative reasoning by asking them to compare two music tracks and answer questions. It probes the model's ability to perform grounded, sentence-level comparative explanations versus simple binary or short-answer discrimination. Use when the user wants to benchmark on Jamendo-MT-QA, or asks about evaluating this task. Reports accuracy, LLM-as-a-Judge score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.09721
  bibtex_key: [REDACTED]
  confidence: high
---

# jamendo-mt-qa-eval

> Jamendo-MT-QA: A Benchmark for Multi-Track Comparative Music Question Answering — Koh et al. (2026) (arXiv:2604.09721, 2026)

## What this evaluates

Evaluates audio-language models on multi-track comparative reasoning by asking them to compare two music tracks and answer questions. It probes the model's ability to perform grounded, sentence-level comparative explanations versus simple binary or short-answer discrimination.

## Datasets

- **Jamendo-MT-QA** — total 36519; splits: full (36519), subset (2010)

## Metrics

- `accuracy` **(primary)** — range: [0, 1]
  - 1 if the predicted answer exactly matches the ground truth, else 0. Applied to yes/no and short-answer questions.
- `BLEU` — range: [0, 1]
  - N-gram precision with brevity penalty to measure surface-form similarity between predicted and gold sentences.
- `ROUGE-1/2/L` — range: [0, 1]
  - Recall-oriented n-gram overlap metrics measuring surface-form similarity.
- `BERTScore` — range: [0, 1]
  - Cosine similarity between contextual embeddings of predicted and gold tokens, averaged across tokens.
- `LLM-as-a-Judge score` **(primary)** — range: [1, 5]
  - 1–5 rating assigned by an LLM based on correctness and comparative soundness using a fixed rubric.

## Input / output format

**Input**: Two audio tracks (or their generated captions) and a comparative question.

**Output**: A text answer: 'yes'/'no', a track identifier, or a natural language sentence.

## Scoring recipe

```python
def score(pred, gold, q_type):
    if q_type == 'yes/no':
        return 1.0 if pred.strip().lower() == gold.strip().lower() else 0.0
    elif q_type == 'short':
        return 1.0 if pred.strip() == gold.strip() else 0.0
    elif q_type == 'sentence':
        bleu = compute_bleu(pred, gold)
        rouge = compute_rouge(pred, gold)
        bert = compute_bertscore_f1(pred, gold)
        llm_score = llm_judge(pred, gold, prompt, rubric) # 1-5
        return {'bleu': bleu, 'rouge': rouge, 'bert_f1': bert, 'llm_judge': llm_score}
```

## Common pitfalls

- Relying solely on BLEU/ROUGE for sentence-level answers underestimates model capability due to low lexical overlap in open-ended generation.
- Comparing caption-based and multi-audio baselines directly without accounting for the intermediate abstraction step can misattribute performance gains.
- LLM-as-a-Judge scores are highly sensitive to the specific prompt and scoring rubric used; results are not directly comparable across different judge configurations.

## Evidence (verbatim from paper)

> For yes/no questions, we report accuracy. For short-answer questions, accuracy is computed by exact match between the predicted track identifier and the ground-truth answer. For sentence-level questions, we report BLEU and ROUGE-1,2,L as surface-form similarity metrics based on n-gram overlap, and BERTScore as a semantic similarity metric.

## Citation

```bibtex
@misc{koh2026jamendomtqa,
  title={Jamendo-MT-QA: A Benchmark for Multi-Track Comparative Music Question Answering},
  author={Koh et al. (2026)},
  year={2026},
  note={arXiv:2604.09721}
}
```

- arXiv: 2604.09721