---
language:
- am
- en
dataset_info:
  features:
  - name: instruction
    dtype: string
  - name: input
    dtype: string
  - name: output
    dtype: string
  - name: prompt_header
    dtype: string
  - name: datasource
    dtype: string
  splits:
  - name: train
    num_bytes: 84855653
    num_examples: 200000
  - name: validation
    num_bytes: 1209980
    num_examples: 1994
  - name: test
    num_bytes: 1306100
    num_examples: 2024
  download_size: 23384531
  dataset_size: 87371733
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
---


# SFT-Data for [Walia-LLM: Enhancing Amharic-LLaMA by Integrating Task-Specific and Generative Datasets](https://arxiv.org/abs/2402.08015)



## Dataset Summary

The Walia dataset is designed to enhance large language models for the Amharic language by:

- Converting existing task-specific datasets (e.g., sentiment analysis, QA, NER) into instruction format.
- Creating new generative datasets (e.g., poem generation, religious lyrics, story generation).
- Translating English instruction datasets (e.g., Alpaca, Dolly) into Amharic for comparative studies.

Each data point follows a structured instruction format with:
- `"instruction"` – a natural language task description,
- `"input"` – optional input text for the task,
- `"output"` – the expected model output in Amharic.

## Supported Tasks

| Task                       | Source/Type       | Notes                      |
|---------------------------|-------------------|----------------------------|
| Sentiment Analysis        | AfriSenti         | 3-class sentiment          |
| Named Entity Recognition  | MasakhaNER        | Personal name extraction   |
| News Classification       | MasakhaNews       | Multilingual topic classes |
| QA                        | AmharicQA         | Wikipedia-based            |
| Summarization             | XL-Sum            | Amharic summaries          |
| Machine Translation       | NLLB, WMT19       | Both directions supported  |
| Poem/Lyrics/Story Gen     | Custom            | Sourced from web/telegram  |
| Spelling Correction       | Synthetic         | Character perturbations    |

## Dataset Structure

```json
{
  "instruction": "Translate the following sentence to Amharic.",
  "input": "Hello, how are you?",
  "output": "ሰላም፣ እንዴት ነህ?"
}
```

## Data Statistics

- ~122,000 instruction samples for training
- ~15,000 for validation and test
- 16+ task types and instruction templates
- All responses are in Amharic (except source text in MT)

## How to Use

You can load the dataset using the Hugging Face `datasets` library:

```python
from datasets import load_dataset

dataset = load_dataset("EthioNLP/walia-amharic-instructions")
print(dataset["train"][0])
```

## Applications

- Supervised fine-tuning (SFT) of LLMs for Amharic
- Cross-lingual instruction tuning experiments
- Evaluation of generative capabilities in low-resource languages

## Related Models

The dataset is used to fine-tune:
- [EthioNLP/Amharic-LLAMA-all-data](https://huggingface.co/EthioNLP/Amharic-LLAMA-all-data)
- [EthioNLP/Amharic_LLAMA_our_data](https://huggingface.co/EthioNLP/Amharic_LLAMA_our_data)

## Citation

Please cite the following paper if you use this dataset:

```bibtex
@inproceedings{azime-etal-2024-walia,
    title = "Walia-{LLM}: Enhancing {A}mharic-{LL}a{MA} by Integrating Task-Specific and Generative Datasets",
    author = "Azime, Israel Abebe  and Tonja, Atnafu Lambebo  and Belay, Tadesse Destaw  and Fuge, Mitiku Yohannes  and Wassie, Aman Kassahun  and Jada, Eyasu Shiferaw  and Chanie, Yonas  and Sewunetie, Walelign Tewabe  and Yimam, Seid Muhie",
    editor = "Al-Onaizan, Yaser  and Bansal, Mohit  and Chen, Yun-Nung",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.findings-emnlp.25/",
    doi = "10.18653/v1/2024.findings-emnlp.25",
    pages = "432--444"
}
```