metadata

base_model: silma-ai/SILMA-9B-Instruct-v1.0
datasets:
  - MohammedNasser/ARabic_Reasoning_QA
language:
  - ar
library_name: transformers
license: apache-2.0
metrics:
  - accuracy
pipeline_tag: question-answering

SILMA-9B-Instruct Fine-Tuned for Arabic Reasoning-QA

This model is a fine-tuned version of silma-ai/SILMA-9B-Instruct-v1.0, optimized for Arabic Question Answering tasks. It excels at providing numerical answers to a wide range of questions in Arabic.

Model Descriptionen

This fine-tuned model is based on the silma-ai/SILMA-9B-Instruct-v1.0 and is designed to answer reasoning questions in Arabic, providing integer-based answers. The model has be fine-tuned using a custom Arabic Reasoning QA dataset, specifically tailored to handle questions ranging from easy to difficult across various topics.

Model Details

Model Name: silma_9b_instruct_ft
Model Type: Language Model
Language: Arabic
Base Model: silma-ai/SILMA-9B-Instruct-v1.0
Fine-Tuning Method: PEFT with LoraConfig
Task: Arabic Question Answering (Numerical Responses)
Training Data: Custom Arabic Reasoning QA dataset
Quantization: 4-bit quantization using bitsandbytes

Features

Optimized for Arabic language understanding and generation
Specialized in providing numerical answers to questions
Efficient inference with 4-bit quantization
Fine-tuned using PEFT with LoraConfig for parameter-efficient training

Training results

Training Loss	Epoch	Step	Validation Loss
2.1356	0.04	10	1.4071
0.8079	0.08	20	0.2825
0.1592	0.12	30	0.1427
0.1202	0.16	40	0.1121
0.1095	0.2	50	0.1071
0.1024	0.24	60	0.1036
0.0993	0.28	70	0.1002
0.091	0.32	80	0.0992
0.1096	0.36	90	0.0965
0.0943	0.4	100	0.0916
0.0882	0.44	110	0.0896
0.0853	0.48	120	0.0848
0.0767	0.52	130	0.0808
0.0778	0.56	140	0.0765
0.0698	0.6	150	0.0734
0.0784	0.64	160	0.0694
0.0648	0.68	170	0.0658
0.0797	0.72	180	0.0630
0.0591	0.76	190	0.0604
0.0557	0.8	200	0.0582
0.0567	0.84	210	0.0561
0.057	0.88	220	0.0534
0.0505	0.92	230	0.0515
0.0483	0.96	240	0.0482
0.0463	1.0	250	0.0463

Training Metrics

Training Loss on wandb 🔗

Usage

Here's a quick example of how to use the model:

from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import torch

model_name = "MohammedNasser/silma_9b_instruct_ft"
user_question = "إذا كان لديك ثلاث سيارات، وبعت واحدة منها، كم سيارة ستبقى لديك؟"

# Load model and tokenizer
import torch
from transformers import pipeline

# Create pipeline
pipe = pipeline(
    "text-generation",
    model=model_name,
    torch_dtype= torch.bfloat16,
    device="cuda",
    return_full_text=False, 
)

messages = [
    {"role": "user", "content": user_question },
]

# Example usage
response = pipe(messages, max_new_tokens=128)
assistant_response = outputs[0]["generated_text"]

print(f"Question: {user_question}")
print(f"Answer: {assistant_response}")

Performance

Our model demonstrates strong performance on Arabic QA tasks, particularly for questions requiring numerical answers. Here are some key metrics:

Eval Loss: 0.046

Limitations

The model is optimized for numerical answers and may not perform as well on open-ended questions.
Performance may vary for dialects or regional variations of Arabic not well-represented in the training data.
The model may occasionally generate incorrect numerical answers for very complex or ambiguous questions.

Fine-tuning Details

The model was fine-tuned using the following configuration:

LoRA Config:
- Alpha: 16
- Dropout: 0.1
- R: 4
Training Hyperparameters:
- Batch Size: 4
- Learning Rate: 2e-4
- Epochs: 3
Hardware: 4 x NVIDIA A100 GPUs

Citation

If you use this model in your research, please cite:

@misc {gaber_2024,
    author       = { {Gaber} },
    title        = { silma_9b_instruct_ft (Revision e54c562) },
    year         = 2024,
    url          = { https://huggingface.co/MohammedNasser/silma_9b_instruct_ft },
    doi          = { 10.57967/hf/3032 },
    publisher    = { Hugging Face }
}

Made with ❤️ by [M. N. Gaber/aiNarabic]