README.md

---
language: en
tags:
- text-to-sql
- codet5
- fine-tuned
- sql-generation
datasets:
- b-mc2/sql-create-context
---

# CodeT5+ 220M — SQL Query Generator

Fine-tuned CodeT5+ 220M on the sql-create-context dataset to generate SQL queries from natural language questions and table schemas.

## Training
- Base model: Salesforce/codet5p-220m (220M parameters)
- Dataset: b-mc2/sql-create-context (55k train / 23k val)
- Method: Full fine-tuning with schema-aware prompting
- Optimizer: AdamW, lr=3e-4, 5 epochs

## Usage
```python
from transformers import T5ForConditionalGeneration, AutoTokenizer
import torch

model = T5ForConditionalGeneration.from_pretrained("your-username/codet5p-sql-generator")
tokenizer = AutoTokenizer.from_pretrained("your-username/codet5p-sql-generator")

context = "CREATE TABLE employees (id INT, name VARCHAR, salary INT)"
question = "What is the average salary?"

src = f"Schema: {context}\nQuestion: {question}"
inputs = tokenizer(src, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=128, num_beams=4)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

## Improvements over GPT-2 baseline
- Schema-aware prompting eliminates column/table hallucination
- Encoder-decoder architecture suited for seq2seq SQL generation
- 10x more training examples (78k vs 7k)
- max_length=512 prevents schema truncation