---
tags:
- Text-to-SQL
- RL
- DPO
license: apache-2.0
base_model:
- mistralai/Mistral-Nemo-Instruct-2407
pipeline_tag: reinforcement-learning
datasets:
- gretelai/synthetic_text_to_sql
---
# Model Card: Mistral-Nemo-Instruct-2407_ORPO- Fine-Tuned for Text-to-SQL

## Model Overview

### Model Name
Mistral-Nemo-Instruct-2407_ORPO

### Base Model
[Mistral-NeMo-Instruct](https://huggingface.co/mistral-nemo-instruct-2407)

### Purpose
This model was fine-tuned to improve accuracy for translating natural language queries into SQL statements, specifically targeting non-technical users. The fine-tuning process compared two methodologies: Direct Preference Optimization (DPO) and Odds Ratio Preference Optimization (ORPO).

---

## Fine-Tuning Methods

### Direct Preference Optimization (DPO)
- A dynamic weight-scaling approach to balance preference alignment and output diversity.
- Uses preference pairs ("selected" vs. "rejected" outputs) to refine model behavior.

### Odds Ratio Preference Optimization (ORPO)
- Leverages binary preference data and an odds ratio-based penalty method.
- Eliminates the need for reward models, offering higher efficiency and scalability.

---

## Dataset

### Training Dataset
- **Source:** Synthetic Text-to-SQL dataset from Gretel AI
- **Size:** 89,495 entries
- **Focus:** Data Query Language (DQL) instructions, complex SQL queries including joins, window functions, and set operations.

### Evaluation Dataset
- **Source:** Mini-Dev dataset from the BIRD benchmark
- **Size:** 500 Text-to-SQL pairs
- **Complexity Levels:** Simple, Medium, Challenging

---

## Evaluation

### Metrics
- **Execution Accuracy (EX):** Percentage of SQL queries executed correctly.

### Results
| Model                   | Execution Accuracy (%) |
|-------------------------|------------------------|
| Mistral-NeMo-Instruct (Base) | Baseline            |
| DPO Fine-Tuned Model    | +0.86%               |
| ORPO Fine-Tuned Model   | +41.38%              |
| ORPO vs. Codestral-22B  | +35.54%              |

![Evaluation Accuracy Comparison](/images/EvaluationAccuracy.png)

---

## Model Use

### Requirements
- Python 3.10+
- PyTorch 2.4+
- CUDA 12.1

### Inference Example
```python
from transformers import pipeline, AutoModelForCausalLM, AutoTokenizer
from peft import PeftConfig,PeftModel

# Load the fine-tuned peft model
peft_config = PeftConfig.from_pretrained("JHuel/Mistral-Nemo-Instruct-2407_DPO_qlora") 
model = AutoModelForCausalLM.from_pretrained(peft_config.base_model_name_or_path)
model = PeftModel.from_pretrained(model, "JHuel/Mistral-Nemo-Instruct-2407_DPO_qlora")


# Load the fine-tuned model
tokenizer = AutoTokenizer.from_pretrained("your-model-name")
model = AutoModelForCausalLM.from_pretrained("your-model-name")
tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-Nemo-Instruct-2407")

# Input a natural language query
response = chatbot(messages)[0]['generated_text']

print(response)
```

---

## Limitations
- The model may not handle queries involving highly specialized or domain-specific SQL operations.
- Training data was limited to synthetic datasets; real-world performance may vary.

---

## Ethical Considerations
- **Bias:** The training dataset was synthetic and may not fully represent real-world linguistic diversity.
- **Misuse:** The model is intended for assisting in SQL generation and should not be used for tasks requiring high levels of security or privacy without additional safeguards.

---

## Citation
If you use this model in your research or applications, please cite:
```
@article{JHuelsEKeuchel,
  title={Evaluation of Fine-Tuning Methods: DPO and ORPO for Text-to-SQL},
  author={Jonathan Hüls and Elina Keuchel.},
  year={2025}
}
```

---

## License
The model is released under the apache-2.0 LICENSE.