Nile-Chat-4B-Transit-Extractor

This model is a fine-tuned version of Nile-Chat-4B designed to extract transit-related entities (Start Points and Destinations) from Egyptian Arabic dialect. It is a core component of the NaviTour project, an AI assistant for Cairo transportation.

Model Description

• Developed by: Hanan Elsaid Elhosary
• Model type: Causal Language Model (Fine-tuned for NER tasks)
• Language(s): Arabic (Egyptian Dialect)
• Finetuned from model: MBZUAI-Paris/Nile-Chat-4B

Intended Uses

This model is specifically optimized to take an Egyptian dialect query about directions and output a structured JSON containing:

• start_point: The location the user is coming from.
• end_point: The intended destination.

Example:

Input: "عايز أروح رمسيس من الدقي" Output: {"start_point": "الدقي", "end_point": "رمسيس"}

Training Details

• Technique: LoRA (Low-Rank Adaptation)
• Quantization: 4-bit (BitsAndBytes)
• Epochs: 1.038 (Stopped early to prevent overfitting)
• Hardware: NVIDIA T4 GPU

Training Data

The model was trained on a custom dataset of 3,762 curated Egyptian Arabic transit queries.

Data Format

Each sample follows the Alpaca instruction format, specifically designed for entity extraction tasks. The dataset includes:

• Instruction-based prompts: Direct and indirect navigation requests.
• Natural Dialect: Real-world Egyptian slang and idioms (e.g., "يا معلم", "أركب إيه").
• Structured Outputs: Clean JSON mapping of transit entities.

Examples:

Instruction	Output (JSON)
"لو عايز أروح جمعية الصناعات الصغيرة في 6 أكتوبر أركب إيه؟"	{"start_point": "", "end_point": "جمعية الصناعات الصغيرة (6 أكتوبر)"}
"يا معلم أروح من كوبري قصر النيل لمستشفى قصر العيني إزاي؟"	{"start_point": "كوبري قصر النيل", "end_point": "مستشفى قصر العيني"}

How to Get Started

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "hananelhosary8/Nile-Chat-4B-Transit-Extractor"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")