---
license: apache-2.0
base_model: Qwen/Qwen2.5-0.5B-Instruct
tags:
- qwen2.5
- lora
- text-to-sql
- sql
- peft
library_name: peft
---

# Qwen2.5-0.5B LoRA Fine-tuned for Text-to-SQL

This is a **LoRA adapter** for [Qwen/Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) fine-tuned on natural language to SQL conversion.

## Quick Links

- 🔗 **Merged Model (Ready-to-use):** [vindows/qwen2.5-0.5b-text-to-sql-merged](https://huggingface.co/vindows/qwen2.5-0.5b-text-to-sql-merged)
- 🖥️ **GGUF for CPU Inference:** [vindows/qwen2.5-0.5b-text-to-sql-gguf](https://huggingface.co/vindows/qwen2.5-0.5b-text-to-sql-gguf)

## Model Performance

### Training Metrics (54 test examples)

| Metric | Base Model | Fine-tuned | Improvement |
|--------|-----------|------------|-------------|
| Loss | 2.1429 | 0.5823 | 72.83% ⬆️ |
| Perplexity | 8.5244 | 1.7901 | 79.00% ⬆️ |

### Spider Benchmark Results (200 examples)

| Metric | Score |
|--------|-------|
| Exact Match | 0.00% |
| Normalized Match | 0.00% |
| Component Accuracy | 91.94% |
| Average Similarity | 21.78% |

**Note:** The model shows strong component understanding but tends to append explanatory text after SQL queries, affecting exact match scores. See limitations below.

## Training Details

- **Training Time:** 2 minutes 5 seconds
- **Epochs:** 3
- **LoRA Rank:** 16
- **LoRA Alpha:** 32
- **Learning Rate:** 2e-4
- **Dataset:** 425 training examples, 54 validation, 54 test

## Usage

### With PEFT (Recommended for fine-tuning/adapters)

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

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-0.5B-Instruct",
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "vindows/qwen2.5-0.5b-text-to-sql")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct", trust_remote_code=True)

# Generate SQL
prompt = "Convert the following natural language question to SQL:\n\nDatabase: concert_singer\nQuestion: How many singers do we have?\n\nSQL:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.1)
sql = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(sql)
```

### Using the Merged Model (Easier)

For easier usage without loading base + adapter separately, use the merged model:

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "vindows/qwen2.5-0.5b-text-to-sql-merged",
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("vindows/qwen2.5-0.5b-text-to-sql-merged")
```

## Limitations

1. **Appends Explanatory Text:** The model tends to append explanatory text or context after generating SQL queries. Post-processing to extract only the SQL statement is recommended.
2. **Hallucinated Table Names (0.5B):** The smaller 0.5B model sometimes invents table names not present in the schema.
3. **Training Data Distribution:** Best performance on queries similar to training examples.

## Recommended Post-Processing

```python
def extract_sql(generated_text):
    # Extract SQL after the "SQL:" marker
    if "SQL:" in generated_text:
        sql = generated_text.split("SQL:")[-1].strip()
    else:
        sql = generated_text

    # Take only the first SQL statement (before extra text)
    if '\n\n' in sql:
        sql = sql.split('\n\n')[0].strip()

    # Remove trailing semicolon if present
    sql = sql.rstrip(';').strip()

    return sql
```

## Files Included

- `adapter_config.json` - LoRA configuration
- `adapter_model.safetensors` - LoRA weights
- `README.md` - This file

## Citation

```bibtex
@misc{qwen2.5-0.5b-text-to-sql,
  title = {Qwen2.5-0.5B LoRA Fine-tuned for Text-to-SQL},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/vindows/qwen2.5-0.5b-text-to-sql}
}
```

## License

Apache 2.0 (inherits from base Qwen2.5 model)