README.md

---
base_model: unsloth/gemma-2b-it-bnb-4bit
tags:
- text-generation-inference
- text-to-sql
- sql
- gemma
- unsloth
- lora
- peft
license: gemma
language:
- en
---

# Gemma SQL Copilot (LoRA)

This is a fine-tuned version of Google's **Gemma 2B-IT** model designed specifically for **Text-to-SQL** generation. It translates natural English instructions into properly formatted SQL queries. 

The model was fine-tuned using [Unsloth](https://github.com/unslothai/unsloth) for efficient 4-bit quantization and LoRA (Low-Rank Adaptation), meaning it is highly memory efficient and can be run locally on consumer GPUs (like an RTX 3050 6GB) with minimal VRAM.

## 🛠️ Intended Use
- **Task:** Natural Language to SQL (Text-to-SQL)
- **Use Case:** Helping data analysts, developers, and business users query databases simply by asking questions in plain English.
- **Environment:** Designed for fast, low-memory inference using `unsloth`.

## ✍️ Prompt Format
This model was trained on a specific prompt structure. To get the best results, you **must** wrap your question in the following format:

```text
### Instruction:
Write a SQL query to find all users who signed up in 2023.

### Response:
<leave this blank for the model to generate the SQL>

💻 Example Usage

# pip install unsloth
from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Noor201/gemma-sql-copilot-lora",
    max_seq_length = 2048,
    dtype = None,
    load_in_4bit = True, 
)
FastLanguageModel.for_inference(model)

prompt = """### Instruction:
Write a SQL query to find the names of all employees in the 'Sales' department who earn more than 50000.

### Response:
"""

inputs = tokenizer([prompt], return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=128)

print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])

⚙️ Training Details

Base Model: google/gemma-2b-it
Training Framework: Unsloth (PEFT/LoRA)
Precision: 4-bit (QLoRA)
Hardware: Trained on a single NVIDIA T4 GPU via Google Colab.

## 📊 Training Results
During the fine-tuning process, the model achieved the following performance metrics on the dataset:
- **Final Training Loss:** 0.0006
- **Final Validation Loss:** 9.3803
- **Epochs:** 2