Base model: LLaMA-2
Fine-tuning method: Supervised fine-tuning (SFT) using LoRA
Language: Amharic
Tasks:
- Sentiment analysis
- Question answering
- Named entity recognition
- News classification
- Summarization
- Machine translation
- Poem/story/lyrics generation
- Spelling correction

Training Data

The model was trained on a custom instruction dataset derived from:

Existing NLP benchmarks (e.g., AfriSenti, AmharicQA, MasakhaNER, MasakhaNews, XL-Sum)
Manually collected generative datasets (e.g., religious lyrics, stories, poems)
Translated instruction datasets (e.g., Alpaca, Dolly)

See EthioNLP/walia-amharic-instructions for the dataset used.

Intended Use

This model is intended for:

Research on instruction tuning in low-resource languages
Generative NLP tasks in Amharic
Evaluating multilingual LLM capabilities

Limitations

Some generative outputs may be verbose or imprecise.
Limited understanding of highly specific Amharic poetic or lyrical structures.
Spell correction and NER performance is still under exploration.

Example Usage

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("israel/LLAMA-Walia-II")
tokenizer = AutoTokenizer.from_pretrained("israel/LLAMA-Walia-II")

prompt = "ስለ አማርኛ ቋንቋ መግለጫ አቅርብ።"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citation

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"
}