README.md

---
license: mit
language:
- yue
tags:
- cantonese
- multimodal
- speech
- benchmark
pretty_name: Multimodal Yue Benchmark (audio + text)
size_categories:
- 10K<n<100K
dataset_info:
- config_name: gsm8k_hiugaai
  features:
  - name: conversations
    list:
    - name: from
      dtype: string
    - name: value
      dtype: string
  - name: audio
    dtype: audio
  splits:
  - name: train
    num_bytes: 173071396.88
    num_examples: 1249
  - name: test
    num_bytes: 9293462.0
    num_examples: 70
  download_size: 165211971
  dataset_size: 182364858.88
- config_name: gsm8k_hiumaan
  features:
  - name: conversations
    list:
    - name: from
      dtype: string
    - name: value
      dtype: string
  - name: audio
    dtype: audio
  splits:
  - name: train
    num_bytes: 139023756.8
    num_examples: 1249
  - name: test
    num_bytes: 7452854.0
    num_examples: 70
  download_size: 128833254
  dataset_size: 146476610.8
- config_name: gsm8k_wanlung
  features:
  - name: conversations
    list:
    - name: from
      dtype: string
    - name: value
      dtype: string
  - name: audio
    dtype: audio
  splits:
  - name: train
    num_bytes: 136425017.456
    num_examples: 1249
  - name: test
    num_bytes: 7371638.0
    num_examples: 70
  download_size: 124585085
  dataset_size: 143796655.456
- config_name: mmlu_hiugaai
  features:
  - name: conversations
    list:
    - name: from
      dtype: string
    - name: value
      dtype: string
  - name: audio
    dtype: audio
  splits:
  - name: train
    num_bytes: 514573984.064
    num_examples: 3516
  - name: test
    num_bytes: 25366480.0
    num_examples: 180
  download_size: 460645039
  dataset_size: 539940464.064
- config_name: mmlu_hiumaan
  features:
  - name: conversations
    list:
    - name: from
      dtype: string
    - name: value
      dtype: string
  - name: audio
    dtype: audio
  splits:
  - name: train
    num_bytes: 425052626.328
    num_examples: 3538
  - name: test
    num_bytes: 21197697.0
    num_examples: 182
  download_size: 369799127
  dataset_size: 446250323.328
- config_name: mmlu_wanlung
  features:
  - name: conversations
    list:
    - name: from
      dtype: string
    - name: value
      dtype: string
  - name: audio
    dtype: audio
  splits:
  - name: train
    num_bytes: 408528411.48
    num_examples: 3538
  - name: test
    num_bytes: 20642289.0
    num_examples: 182
  download_size: 350137445
  dataset_size: 429170700.48
configs:
- config_name: gsm8k_hiugaai
  data_files:
  - split: train
    path: gsm8k_hiugaai/train-*
  - split: test
    path: gsm8k_hiugaai/test-*
- config_name: gsm8k_hiumaan
  data_files:
  - split: train
    path: gsm8k_hiumaan/train-*
  - split: test
    path: gsm8k_hiumaan/test-*
- config_name: gsm8k_wanlung
  data_files:
  - split: train
    path: gsm8k_wanlung/train-*
  - split: test
    path: gsm8k_wanlung/test-*
- config_name: mmlu_hiugaai
  data_files:
  - split: train
    path: mmlu_hiugaai/train-*
  - split: test
    path: mmlu_hiugaai/test-*
  default: true
- config_name: mmlu_hiumaan
  data_files:
  - split: train
    path: mmlu_hiumaan/train-*
  - split: test
    path: mmlu_hiumaan/test-*
- config_name: mmlu_wanlung
  data_files:
  - split: train
    path: mmlu_wanlung/train-*
  - split: test
    path: mmlu_wanlung/test-*
---

# Multimodal Yue Benchmark

Cantonese **audio + text** benchmark derived from **[BillBao/Yue-Benchmark](https://huggingface.co/datasets/BillBao/Yue-Benchmark)** (Yue-GSM8K & Yue-MMLU-style tasks). We kept the same task content in Cantonese and added **TTS** for three Cantonese speakers (`hiugaai`, `hiumaan`, `wanlung`).

## Subsets & splits

| Config name | Task | Speaker | Splits |
|-------------|------|---------|--------|
| `mmlu_*` | multiple-choice (MMLU-style) | per speaker | `train`, `test` |
| `gsm8k_*` | math word problems (GSM8K-style) | per speaker | `train`, `test` |

Example:

```python
from datasets import load_dataset

ds = load_dataset("J017athan/Multimodal-Yue-Benchmark", "mmlu_hiugaai", split="train")
```

## Format

Each row has **`conversations`** (ShareGPT-style list of `{"from": "human"|"gpt", "value": "..."}`) and **`audio`** (embedded in Parquet when using the default uploader). Compatible with LLaMA-Factory.

## Upstream

Original **[Yue-Benchmark](https://huggingface.co/datasets/BillBao/Yue-Benchmark)**.