hamza-amin/alif-emergency-finetuned

Alif-1.0-8B Emergency Response (Urdu)

Fine-tuned version of Alif-1.0-8B-Instruct for Urdu emergency call handling.

🚀 Try the Model in Colab

Model Description

This model is an Urdu conversational language model fine-tuned to understand and respond to emergency call scenarios.
It is adapted from Alif-1.0-8B-Instruct using QLoRA on a curated Urdu emergency calls dataset (~5k samples).

• Base model: large-traversaal/Alif-1.0-8B-Instruct
• Language: Urdu
• Task: Emergency call understanding and response
• Fine-tuning: QLoRA (4-bit NF4, LoRA adapters)

---

Intended Use

Intended:

• Research on Urdu emergency dialogue systems
• Prototyping and academic experimentation
• Conversational AI demos

Not intended:

• Real-world emergency response
• Medical, legal, or safety-critical decision-making
• Autonomous deployment without human supervision

---

Training Data

Fine-tuned on the Urdu Emergency Calls Dataset:
https://huggingface.co/datasets/hamza-amin/urdu-emergency-calls

The dataset contains instruction-style emergency conversations in Urdu.

---

Training Details

• LoRA rank: 8
• LoRA alpha: 16
• Target modules: q_proj, v_proj
• Epochs: 1
• Learning rate: 2e-4
• Precision: bf16
• Optimizer: paged_adamw_8bit

---

Evaluation

• Validation loss: ~0.39
• Perplexity: ~1.5

These results indicate stable training and effective domain adaptation for a small, task-specific dataset.

---

Limitations

• Trained on a limited, curated dataset
• May fail on ambiguous or unseen emergency scenarios
• Not suitable for real emergency use without human oversight

---

Quick Start

from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel
import torch

base_model_name = "large-traversaal/Alif-1.0-8B-Instruct"
adapter_name = "hamza-amin/alif-emergency-finetuned"

tokenizer = AutoTokenizer.from_pretrained(base_model_name, trust_remote_code=True)

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=False,
)

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True
)

model = PeftModel.from_pretrained(base_model, adapter_name)

Example Outputs

Query: میں ملتان کے شاہ رکن عالم کالونی میں ہوں، ایک شخص کنویں میں گر گیا ہے۔
Response: کیا آپ صحیح مقام بتا سکتے ہیں؟

Query: جی ہاں، میں نے بس حادثے میں زخمیوں کی مدد شروع کر دی ہے۔
Response: ایمبولینس روانہ ہو چکی ہے، آپ زخمیوں کا خیال رکھیں۔

Citation

@misc{alif-emergency-urdu,
  author = {Hamza Amin},
  title = {Alif Emergency Response Model for Urdu},
  year = {2024},
  publisher = {HuggingFace},
  url = {https://huggingface.co/hamza-amin/alif-emergency-finetuned}
}

Acknowledgments

• Base model: Alif-1.0-8B-Instruct
• Dataset: urdu-emergency-calls
• Training platform: Kaggle