File size: 4,984 Bytes
c19331a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
---
license: mit
language:
- ar
base_model:
- Qwen/Qwen2.5-1.5B-Instruct
pipeline_tag: text2text-generation
library_name: transformers
tags:
- Text-To-SQL
- Arabic
- Spider
- SQL
---

# Model Card for Arabic Text-To-SQL (OsamaMo)

## Model Details

### Model Description

This model is fine-tuned on the Spider dataset with Arabic-translated questions for the Text-To-SQL task. It is based on **Qwen/Qwen2.5-1.5B-Instruct** and trained using LoRA on Kaggle for 15 hours on a **P100 8GB GPU**.

- **Developed by:** Osama Mohamed ([OsamaMo](https://huggingface.co/OsamaMo))
- **Funded by:** Self-funded
- **Shared by:** Osama Mohamed
- **Model type:** Text-to-SQL fine-tuned model
- **Language(s):** Arabic (ar)
- **License:** MIT
- **Finetuned from:** Qwen/Qwen2.5-1.5B-Instruct

### Model Sources

- **Repository:** [Hugging Face Model Hub](https://huggingface.co/OsamaMo/Arabic_Text-To-SQL)
- **Dataset:** Spider (translated to Arabic)
- **Training Script:** [LLaMA-Factory](https://github.com/huggingface/transformers/tree/main/src/transformers/models/llama_factory)

## Uses

### Direct Use

This model is intended for converting **Arabic natural language questions** into SQL queries. It can be used for database querying in Arabic-speaking applications.

### Downstream Use

Can be fine-tuned further for specific databases or Arabic dialect adaptations.

### Out-of-Scope Use

- The model is **not** intended for direct execution of SQL queries.
- Not recommended for non-database-related NLP tasks.

## Bias, Risks, and Limitations

- The model might generate incorrect or non-optimized SQL queries.
- Bias may exist due to dataset translations and model pretraining data.

### Recommendations

- Validate generated SQL queries before execution.
- Ensure compatibility with specific database schemas.

## How to Get Started with the Model

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"
base_model_id = "Qwen/Qwen2.5-1.5B-Instruct"
finetuned_model_id = "OsamaMo/Arabic_Text-To-SQL_using_Qwen2.5-1.5B"

model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16
)

model.load_adapter(finetuned_model_id)

tokenizer = AutoTokenizer.from_pretrained(base_model_id)

def generate_resp(messages):
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )

    model_inputs = tokenizer([text], return_tensors="pt").to(device)

    generated_ids = model.generate(
        model_inputs.input_ids,
        max_new_tokens=1024,
        do_sample=False, top_k=None, temperature=None, top_p=None,
    )

    generated_ids = [
        output_ids[len(input_ids):]
        for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
    ]

    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

    return response
```

## Training Details

### Training Data

- Dataset: **Spider (translated into Arabic)**
- Preprocessing: Questions converted to Arabic while keeping SQL queries unchanged.
- Training format:
  - System instruction guiding Arabic-to-SQL conversion.
  - Database schema provided for context.
  - Arabic user queries mapped to correct SQL output.
  - Output is strictly formatted SQL queries enclosed in markdown code blocks.

### Training Procedure

#### Training Hyperparameters

- **Batch size:** 1 (per device)
- **Gradient accumulation:** 4 steps
- **Learning rate:** 1.0e-4
- **Epochs:** 3
- **Scheduler:** Cosine
- **Warmup ratio:** 0.1
- **Precision:** bf16

#### Speeds, Sizes, Times

- **Training time:** 15 hours on **NVIDIA P100 8GB**
- **Checkpointing every:** 500 steps

## Evaluation

### Testing Data

- Validation dataset: Spider validation set (translated to Arabic)

### Metrics

- Exact Match (EM) for SQL correctness
- Execution Accuracy (EX) on databases

### Results

- Model achieved **competitive SQL generation accuracy** for Arabic queries.
- Further testing required for robustness.

## Environmental Impact

- **Hardware Type:** NVIDIA Tesla P100 8GB
- **Hours used:** 15
- **Cloud Provider:** Kaggle
- **Carbon Emitted:** Estimated using [ML Impact Calculator](https://mlco2.github.io/impact#compute)

## Technical Specifications

### Model Architecture and Objective

- Transformer-based **Qwen2.5-1.5B** architecture.
- Fine-tuned for Text-to-SQL task using LoRA.

### Compute Infrastructure

- **Hardware:** Kaggle P100 GPU (8GB VRAM)
- **Software:** Python, Transformers, LLaMA-Factory, Hugging Face Hub

## Citation

If you use this model, please cite:

```bibtex
@misc{OsamaMo_ArabicSQL,
  author = {Osama Mohamed},
  title = {Arabic Text-To-SQL Model},
  year = {2024},
  howpublished = {\url{https://huggingface.co/OsamaMo/Arabic_Text-To-SQL}}
}
```

## Model Card Contact

For questions, contact **Osama Mohamed** via Hugging Face ([OsamaMo](https://huggingface.co/OsamaMo)).