| --- |
| language: |
| - en |
| license: llama3.2 |
| base_model: unsloth/Llama-3.2-3B |
| tags: |
| - text-generation |
| - sql |
| - distributed-databases |
| - qlora |
| - peft |
| - fine-tuned |
| - e-commerce |
| pipeline_tag: text-generation |
| --- |
| |
| # Llama 3.2 3B — E-commerce Distributed SQL |
|
|
| Fine-tuned version of Llama 3.2 3B that converts natural language questions |
| into SQL queries for distributed e-commerce databases. |
|
|
| ## Example |
|
|
| **Input:** |
| ``` |
| ### Instruction: |
| Convert to distributed SQL |
| |
| ### Input: |
| Find all customers who spent more than 1000 euros in Germany |
| |
| ### Response: |
| ``` |
|
|
| **Output:** |
| ```sql |
| SELECT * FROM customers |
| WHERE country = 'Germany' AND amount > 1000; |
| ``` |
|
|
| ## Model Details |
|
|
| | Property | Value | |
| |----------|-------| |
| | Base model | Llama 3.2 3B | |
| | Fine-tuning method | QLoRA (4-bit quantization + LoRA) | |
| | LoRA rank | 16 | |
| | Trainable parameters | 0.14% | |
| | Training GPU | Google Colab T4 (free tier) | |
| | Training time | ~20 minutes | |
| | Dataset size | 25 examples | |
| | Training epochs | 3 | |
|
|
| ## Training Details |
|
|
| Fine-tuned using QLoRA — 4-bit NF4 quantization with LoRA adapters on the |
| attention layers (`q_proj`, `v_proj`). This reduced memory requirements enough |
| to train on a free Colab T4 GPU (15GB VRAM) in under 20 minutes, while only |
| updating 0.14% of parameters. |
|
|
| **Libraries used:** HuggingFace Transformers, PEFT, TRL (SFTTrainer), |
| bitsandbytes, datasets |
|
|
| ## Dataset |
|
|
| 25 natural language → SQL pairs covering distributed e-commerce scenarios: |
| - Orders across regions and shards |
| - Inventory across warehouses |
| - Customer analytics and segmentation |
| - Revenue aggregations |
| - JOIN queries across fragmented tables |
|
|
| **Prompt format used during training:** |
| ``` |
| ### Instruction: |
| Convert to distributed SQL |
| |
| ### Input: |
| {natural language question} |
| |
| ### Response: |
| {SQL query} |
| ``` |
|
|
| ## How to Use |
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline |
| from peft import PeftModel |
| |
| # Load base model + adapter |
| base = AutoModelForCausalLM.from_pretrained("unsloth/Llama-3.2-3B") |
| model = PeftModel.from_pretrained(base, "haricharanhl22/ecommerce-distributed-sql") |
| tokenizer = AutoTokenizer.from_pretrained("haricharanhl22/ecommerce-distributed-sql") |
| |
| pipe = pipeline("text-generation", model=model, tokenizer=tokenizer) |
| |
| query = """### Instruction: |
| Convert to distributed SQL |
| |
| ### Input: |
| Find top 5 customers by total order value |
| |
| ### Response:""" |
| |
| result = pipe(query, max_new_tokens=100, do_sample=False) |
| print(result[0]["generated_text"]) |
| ``` |
|
|
| ## Limitations |
|
|
| - Trained on a small dataset (25 examples) — works best for common query patterns |
| - Optimized for e-commerce schemas (orders, customers, products, inventory) |
| - May not generalize well to very complex multi-level nested subqueries |
| - SQL dialect closest to standard SQL / SQLite |
|
|
| ## Author |
|
|
| **Hari Charan Hosakote Lokesh** |
| M.Sc. Digital Engineering — Otto-von-Guericke-Universität Magdeburg |
|
|
| - GitHub: [haricharanhl22](https://github.com/haricharanhl22) |
| - LinkedIn: [haricharanhl22](https://linkedin.com/in/haricharanhl22) |
| - Live project: [ai-bewerbung-assistant.vercel.app](https://ai-bewerbung-assistant.vercel.app) |
