Llama-3.1-8b-Instruct-Egy-Arabic

This finetune was a failure, dont count on it being functional at all

wanted to see how hard it'd be to fine-tune an LLM on the way me and my friends actually talk, which is a mix of Egyptian Arabic, English, and transliterated Arabic (Latin script with numbers for Arabic letters).

What it can do

• Respond casually the way you'd talk in a group chat
• Handle the transliteration system (3, 7, 2, etc.) well enough.

Honest limitations

• it can't hold a full length conversation in Egyptian colloquial Arabic, English, or a mix of both
• Training was cut short — this is not a finished model
• It may randomly switch languages mid-response
• Output quality is inconsistent; it can say something coherent then immediately say something off
• Small chance of producing slurs or offensive content since the training data was unfiltered real conversations
• Not suitable for any use

Training details

• Base model: unsloth/meta-llama-3.1-8b-instruct-unsloth-bnb-4bit
• Method: QLoRA fine-tuning via Unsloth
• Hardware: Google Colab free tier (T4 GPU)
• Developed by: MatteioAI
• License: Apache-2.0

Trained 2x faster with Unsloth

<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>