---
license: mit
task_categories:
- text-generation
language:
- zh
- en
tags:
- benchmark
- podcast
- instruction-following
- long-context
- script-generation
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/podbench_800.jsonl
---

# PodBench

A benchmark for **instruction-aware, audio-oriented podcast script generation**.

📃 [Paper (ACL 2026)](https://aclanthology.org/2026.acl-long.2019/) &nbsp;•&nbsp;
💻 [Code & evaluation toolkit](https://github.com/xucncn/PodBench)

PodBench evaluates whether a model can turn real-world source material into a
podcast script that satisfies an explicit user brief. It contains **800 queries**
(400 Chinese, 400 English) with inputs up to **21K tokens**, spanning eight
requirement dimensions including multi-speaker instructions.

## Quick Start

```python
from datasets import load_dataset

ds = load_dataset("cnxu/PodBench", split="test")

# `input_prompt` is the fully assembled model input.
print(ds[0]["input_prompt"])
```

Generate a script for each `input_prompt`, then score the outputs with the
toolkit in the [GitHub repository](https://github.com/xucncn/PodBench):

```bash
export OPENROUTER_API_KEY="sk-or-v1-your-key-here"
python evaluate_benchmark.py --model_file generations.jsonl --output_file ./results/my_model_eval.jsonl
python calculate_scores.py --results_dir ./results --detailed
```

## Fields

| Field | Type | Description |
|---|---|---|
| `id` | string | Sample identifier |
| `instruction` | string | The user requirement under test |
| `input_prompt` | string | **Fully assembled model input** — feed this to your model |
| `meta` | string (JSON) | Everything else: source material, speaker spec, dimensions, languages, token counts, provenance |

## Evaluation

Scripts are scored on two axes, averaged into the final metric:

1. **Instruction Following (0–100)** — the judge derives a sample-specific
   checklist from the instruction and verifies each item against the script.
2. **Podcast Script Quality (0–100)** — an audio-first rubric covering Content
   Substance (45), Narrative Engagement (30), and Conversational Naturalness (25).

Both rubrics and the scoring scripts are in the
[GitHub repository](https://github.com/xucncn/PodBench). Results reported in the
paper use `anthropic/claude-opus-4.5` as the judge.

## Citation

```bibtex
@inproceedings{xu-etal-2026-podbench,
    title = "{P}od{B}ench: A Comprehensive Benchmark for Instruction-Aware Audio-Oriented Podcast Script Generation",
    author = "Xu, Chenning  and
      Zheng, Mao  and
      Zheng, Mingyu  and
      Song, Mingyang",
    booktitle = "Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
    month = jul,
    year = "2026",
    address = "San Diego, California, United States",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.acl-long.2019/",
    doi = "10.18653/v1/2026.acl-long.2019",
    pages = "43605--43621",
    ISBN = "979-8-89176-390-6"
}
```