74 lines
4.4 KiB
Markdown
74 lines
4.4 KiB
Markdown
|
---
|
||
|
|
license: mit
|
||
|
|
---
|
||
|
|
```markdown
|
||
|
|
# Llama-2 7B Text-to-SQL (Fine-Tuned with LoRA)
|
||
|
|
|
||
|
|
## 📌 Model Overview
|
||
|
|
This repository contains a fine-tuned version of Meta's **Llama-3 (8B)**, optimized specifically for the **Natural Language to SQL (Text-to-SQL)** generation task. By converting conversational English into executable SQL queries, this model is designed to bridge the gap between non-technical stakeholders and complex relational databases.
|
||
|
|
|
||
|
|
- **Developer:** DanielMartin Arogyasami
|
||
|
|
- **Base Model:** Meta-Llama-3-8B
|
||
|
|
- **Task:** Text-to-SQL (Code Generation)
|
||
|
|
- **Fine-Tuning Methodology:** Low-Rank Adaptation (LoRA) / PEFT
|
||
|
|
- **Language:** English, SQL
|
||
|
|
- **License:** Meta Llama 3 Community License / MIT (for fine-tuned weights)
|
||
|
|
|
||
|
|
## 🎯 Intended Use Cases
|
||
|
|
This model is highly specialized for deployment in regulated enterprise environments (e.g., healthcare, finance), where data sovereignty is paramount.
|
||
|
|
- **Enterprise Data Retrieval:** Empowering business users to query databases using natural language, significantly reducing reliance on specialized SQL programmers.
|
||
|
|
- **Agentic AI Workflows:** Serving as the SQL-generation agent within larger Retrieval-Augmented Generation (RAG) and enterprise AI architectures.
|
||
|
|
- **Privacy-Preserving Analytics:** Allowing companies to run text-to-SQL conversions entirely on-premises or within air-gapped Virtual Private Clouds (VPCs). This ensures compliance with HIPAA and FDA CFR Part 11, as no sensitive data is transmitted to external proprietary APIs.
|
||
|
|
|
||
|
|
## ⚙️ Technical Details & Training
|
||
|
|
This model was trained using **Parameter-Efficient Fine-Tuning (PEFT)**. Specifically, **LoRA (Low-Rank Adaptation)** was applied to the foundational Llama-3 model. This approach adapts the foundational reasoning capabilities of Llama-3 to the strict syntax requirements of SQL generation, while significantly reducing computational overhead.
|
||
|
|
|
||
|
|
* **Adapter:** LoRA
|
||
|
|
* **Target Modules:** Attention weights (`q_proj`, `v_proj`)
|
||
|
|
* **Precision:** FP16 / 4-bit Quantization (QLoRA) supported for edge-deployment.
|
||
|
|
* **Architecture:** Auto-Regressive Transformer.
|
||
|
|
|
||
|
|
## 🚀 How to Use (Inference)
|
||
|
|
You can load this model and run inference using the `transformers` and `peft` libraries from Hugging Face.
|
||
|
|
|
||
|
|
```python
|
||
|
|
import torch
|
||
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||
|
|
from peft import PeftModel
|
||
|
|
|
||
|
|
# Load base model and tokenizer
|
||
|
|
base_model_id = "meta-llama/Meta-Llama-3-8B"
|
||
|
|
adapter_id = "Arogyasami/Llama-3-8b-text2sql-finetune"
|
||
|
|
|
||
|
|
tokenizer = AutoTokenizer.from_pretrained(base_model_id)
|
||
|
|
base_model = AutoModelForCausalLM.from_pretrained(base_model_id, torch_dtype=torch.float16, device_map="auto")
|
||
|
|
|
||
|
|
# Load LoRA adapter
|
||
|
|
model = PeftModel.from_pretrained(base_model, adapter_id)
|
||
|
|
|
||
|
|
# Define your schema and question
|
||
|
|
schema = "CREATE TABLE Employees (ID int, Name varchar(255), Department varchar(255), Salary int);"
|
||
|
|
question = "What is the average salary of employees in the Sales department?"
|
||
|
|
|
||
|
|
prompt = f"Given the following database schema:\n{schema}\n\nWrite a SQL query to answer this question: {question}\n\nSQL:"
|
||
|
|
|
||
|
|
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
|
||
|
|
|
||
|
|
# Generate SQL
|
||
|
|
outputs = model.generate(**inputs, max_new_tokens=100)
|
||
|
|
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
||
|
|
```
|
||
|
|
|
||
|
|
## ⚠️ Limitations & Bias
|
||
|
|
- **Deterministic Requirement:** Generative AI models can "hallucinate." While fine-tuned to minimize this, generated queries must be validated before execution, especially in clinical contexts where erroneous data retrieval could impact regulatory submissions.
|
||
|
|
- **Schema Complexity:** While the model performs exceptionally well on standard schemas, highly complex schemas with dozens of nested joins may reduce accuracy.
|
||
|
|
- **Security:** Always execute generated SQL in a safe, read-only environment. The model does not inherently enforce database permissions.
|
||
|
|
|
||
|
|
## 🌍 Environmental Impact
|
||
|
|
By utilizing LoRA instead of full-parameter fine-tuning, the computational cost and carbon footprint required to train this model were drastically reduced. Its 8B parameter size allows it to be served efficiently on consumer-grade hardware (e.g., NVIDIA RTX 4090 or A10G), democratizing advanced AI access.
|
||
|
|
|
||
|
|
## 🤝 Acknowledgements & Citation
|
||
|
|
- Meta for the foundational Llama-3 architecture.
|
||
|
|
- If this model assists in your research or enterprise architecture, please cite:
|
||
|
|
Arogyasami, DanielMartin. (2024). *Llama-3 8B Text-to-SQL (Fine-Tuned with LoRA)*. Hugging Face.
|
||
|
|
```
|