---
dataset_info:
- config_name: default
  features:
  - name: id
    dtype: string
  - name: label
    dtype: string
  - name: label_text
    dtype: string
  - name: text
    dtype: string
  - name: lang
    dtype: string
  splits:
  - name: train
    num_bytes: 1056214
    num_examples: 11514
  - name: test
    num_bytes: 271925
    num_examples: 2974
  - name: validation
    num_bytes: 185181
    num_examples: 2033
  download_size: 519296
  dataset_size: 1513320
- config_name: intent_with_options
  features:
  - name: id
    dtype: string
  - name: label
    dtype: string
  - name: label_text
    dtype: string
  - name: text
    dtype: string
  - name: lang
    dtype: string
  - name: options
    list:
    - name: label
      dtype: string
    - name: label_text
      dtype: string
  splits:
  - name: train
    num_bytes: 3305760
    num_examples: 2974
  download_size: 2966668
  dataset_size: 3305760
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: validation
    path: data/validation-*
- config_name: intent_with_options
  data_files:
  - split: train
    path: intent_with_options/train-*
---
# Turkish Intent Routing Dataset

This dataset is based on the Turkish subset of [`mteb/amazon_massive_intent`](https://huggingface.co/datasets/mteb/amazon_massive_intent).

## `intent_with_options` Configuration

The `intent_with_options` configuration was created for multiple-choice intent routing evaluation.

- Added Turkish descriptions (`label_text`) for each intent.
- Used [`alibayram/embeddingmagibu-200m`](https://huggingface.co/alibayram/embeddingmagibu-200m) to calculate semantic similarity between user utterances and intent descriptions.
- Selected the 9 most similar incorrect intents as distractors for each example.
- Included the correct intent, resulting in 10 options per example.
- Randomly shuffled the options

## Usage

```python
from datasets import load_dataset

dataset = load_dataset(
    "canbingol/amazon_massive_intent_tr",
    "intent_with_options"
)
```

This configuration is intended for evaluating how well models distinguish between semantically similar intents in Turkish.