---
license: mit
language:
- ur
- en
task_categories:
- text-classification
- token-classification
pretty_name: bilingual-ticket-triage-dataset
tags:
- synthetic
- support-tickets
- roman-urdu
- entity-extraction
---

# bilingual-ticket-triage-dataset

Fully **synthetic** support-ticket dataset used to fine-tune [bilingual-ticket-triage-adapter](https://huggingface.co/abuzarkhan/bilingual-ticket-triage-adapter) (QLoRA, Qwen2.5-3B-Instruct). Tickets are written in **Roman Urdu**, **Urdu script**, **English**, and code-mixed variants, mirroring how Pakistani customers actually write support emails.

> **No real customer data.** All personas, names, addresses, and order numbers are fabricated. No real email addresses appear in the source seeds.

## Content

Each record:

```json
{
  "text": "I missed my delivery attempt for order ORD-93811, driver call nahi pick kar raha. Ab dobara kab aayega Smart Thermostat?",
  "category": "shipping_delivery",
  "urgency": "medium",
  "entities": {
    "order_id": "ORD-93811",
    "product_name": "Smart Thermostat",
    "account_email": null,
    "sentiment": "neutral"
  }
}
```

- `category` — one of 8: billing, shipping_delivery, refund_return, technical_issue, account_access, product_complaint, general_inquiry, other
- `urgency` — low / medium / high
- `entities` — `order_id`, `product_name`, `account_email` (null when absent), and `sentiment` (positive / neutral / negative)

## Splits

| Split | Records |
| --- | --- |
| train | 1,414 |
| val | 151 |
| test | 148 |

## Provenance (as-trained note)

The dataset is **kept byte-identical to what the model trained on** so the reported evaluation numbers stay exact. Two things that follow from this:

1. Personas are fabricated, but example addresses may use **real-looking domains** (e.g. `gmail.com`) — this is intentional to match real-world ticket distributions; no such address is a real customer.
2. A `_diversified` flag appears on some records — a data-augmentation artifact from the generation pipeline; it is not a label.

Seeds were written manually per category, then programmatically expanded and diversified (scripts in [support-ticket-automation](https://github.com/abuzarai/support-ticket-automation), see `data/raw/*_seed.txt` and `src/data_prep/`).

## Usage

```python
from datasets import load_dataset
ds = load_dataset("abuzarkhan/bilingual-ticket-triage-dataset")
```

## License

MIT. The model fine-tuned on this data is separately licensed under the [Qwen research license](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct/blob/main/LICENSE).