README.md

---
dataset_info:
- config_name: AnimalQA
  features:
  - name: file
    dtype: string
  - name: audio
    dtype: audio
  - name: attribute_label
    dtype: string
  - name: single_instruction
    dtype: string
  - name: single_answer
    dtype: string
  - name: multi_instruction
    dtype: string
  - name: multi_answer
    dtype: string
  splits:
  - name: kazakh
    num_bytes: 72729461
    num_examples: 500
  - name: russian
    num_bytes: 72769565
    num_examples: 500
  download_size: 144956117
  dataset_size: 145499026
- config_name: EmotionQA
  features:
  - name: file
    dtype: string
  - name: audio
    dtype: audio
  - name: attribute_label
    dtype: string
  - name: single_instruction
    dtype: string
  - name: single_answer
    dtype: string
  - name: multi_instruction
    dtype: string
  - name: multi_answer
    dtype: string
  splits:
  - name: kazakh
    num_bytes: 49480088
    num_examples: 500
  - name: russian
    num_bytes: 49498398
    num_examples: 500
  download_size: 98304628
  dataset_size: 98978486
- config_name: GenderQA
  features:
  - name: file
    dtype: string
  - name: audio
    dtype: audio
  - name: attribute_label
    dtype: string
  - name: single_instruction
    dtype: string
  - name: single_answer
    dtype: string
  - name: multi_instruction
    dtype: string
  - name: multi_answer
    dtype: string
  splits:
  - name: kazakh
    num_bytes: 93143548
    num_examples: 500
  - name: russian
    num_bytes: 93172347
    num_examples: 500
  download_size: 185881785
  dataset_size: 186315895
- config_name: LanguageQA
  features:
  - name: file
    dtype: string
  - name: audio
    dtype: audio
  - name: attribute_label
    dtype: string
  - name: single_instruction
    dtype: string
  - name: single_answer
    dtype: string
  - name: multi_instruction
    dtype: string
  - name: multi_answer
    dtype: string
  splits:
  - name: kazakh
    num_bytes: 92543832
    num_examples: 500
  - name: russian
    num_bytes: 92591526
    num_examples: 500
  download_size: 184646141
  dataset_size: 185135358
configs:
- config_name: AnimalQA
  data_files:
  - split: kazakh
    path: AnimalQA/kazakh-*
  - split: russian
    path: AnimalQA/russian-*
- config_name: EmotionQA
  data_files:
  - split: kazakh
    path: EmotionQA/kazakh-*
  - split: russian
    path: EmotionQA/russian-*
- config_name: GenderQA
  data_files:
  - split: kazakh
    path: GenderQA/kazakh-*
  - split: russian
    path: GenderQA/russian-*
- config_name: LanguageQA
  data_files:
  - split: kazakh
    path: LanguageQA/kazakh-*
  - split: russian
    path: LanguageQA/russian-*
---

# SAKURA — Kazakh & Russian

This dataset is a machine-translated Kazakh and Russian adaptation of the [SAKURA benchmark](https://huggingface.co/SLLM-multi-hop) for evaluating single-hop and multi-hop reasoning in Audio-Language Models.

The original SAKURA benchmark probes whether audio-language models can not only perceive an attribute from speech/audio (single-hop) but also reason over that attribute using world knowledge (multi-hop). This release localizes the textual instructions and reference answers into Kazakh and Russian, enabling evaluation of multilingual audio-language models on the same audio stimuli.

Audio is unchanged. Only the textual fields (`single_instruction`, `single_answer`, `multi_instruction`, `multi_answer`) were machine-translated. The audio recordings are bit-identical copies of the originals.

## Dataset structure

The benchmark covers four tracks, each probing a different perceptual attribute:

| Subset       | Purpose    |
|--------------|----------------------------------|
| `AnimalQA`   | Which animal produces the sound  |
| `EmotionQA`  | Emotion conveyed by the speaker  |
| `GenderQA`   | Gender of the speaker            |
| `LanguageQA` | Language being spoken            |

Each subset contains two splits with parallel content (same audio, translated text):

- `kazakh` — instructions and answers translated into Kazakh
- `russian` — instructions and answers translated into Russian