|
|
--- |
|
|
language: |
|
|
- en |
|
|
tags: |
|
|
- sql |
|
|
- text-to-sql |
|
|
- daraz |
|
|
- llama3 |
|
|
- unsloth |
|
|
- ecommerce |
|
|
license: apache-2.0 |
|
|
datasets: |
|
|
- custom |
|
|
base_model: unsloth/llama-3-8b-bnb-4bit |
|
|
--- |
|
|
|
|
|
# drz-sql-llama3 |
|
|
|
|
|
This model is a fine-tuned version of Llama 3 (8B) for generating SQL queries specific to the Daraz e-commerce platform. |
|
|
|
|
|
## Model Description |
|
|
|
|
|
- **Base Model:** Llama 3 8B (4-bit quantized) |
|
|
- **Fine-tuning Method:** LoRA (Low-Rank Adaptation) |
|
|
- **Training Data:** 20 Daraz-specific SQL query examples |
|
|
- **Use Case:** Converting natural language questions to SQL queries for Daraz analytics |
|
|
|
|
|
## Training Details |
|
|
|
|
|
- **Framework:** Unsloth |
|
|
- **LoRA Rank:** 16 |
|
|
- **Training Steps:** 100 |
|
|
- **Batch Size:** 2 |
|
|
- **Gradient Accumulation:** 4 |
|
|
- **Learning Rate:** 0.0002 |
|
|
|
|
|
## Key Features |
|
|
|
|
|
This model understands Daraz-specific: |
|
|
- Table schemas (e.g., `daraz_cdm.dwd_drz_trd_core_df`, `daraz_cdm.dwd_drz_prd_sku_extension`) |
|
|
- Business logic (Choice classification, KAM assignments, industry mapping) |
|
|
- Query patterns (MAX_PT for partitions, DATEADD for date filtering) |
|
|
- Metrics (GMV, L7/L30 calculations, order types) |
|
|
|
|
|
## Usage |
|
|
|
|
|
```python |
|
|
from unsloth import FastLanguageModel |
|
|
|
|
|
# Load model |
|
|
model, tokenizer = FastLanguageModel.from_pretrained( |
|
|
model_name = "Bilal326/drz-sql-llama3", |
|
|
max_seq_length = 2048, |
|
|
dtype = None, |
|
|
load_in_4bit = True, |
|
|
) |
|
|
|
|
|
FastLanguageModel.for_inference(model) |
|
|
|
|
|
# Generate SQL |
|
|
alpaca_prompt = """Below is an instruction that describes a task, paired with an input that provides further context. Write a response that appropriately completes the request. |
|
|
|
|
|
### Instruction: |
|
|
{} |
|
|
|
|
|
### Input: |
|
|
{} |
|
|
|
|
|
### Response: |
|
|
{}""" |
|
|
|
|
|
prompt = alpaca_prompt.format( |
|
|
"Generate SQL for the following request:", |
|
|
"Get total GMV for last 30 days in Pakistan", |
|
|
"" |
|
|
) |
|
|
|
|
|
inputs = tokenizer([prompt], return_tensors="pt").to("cuda") |
|
|
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.5) |
|
|
print(tokenizer.decode(outputs[0])) |
|
|
``` |
|
|
|
|
|
## Example Queries |
|
|
|
|
|
The model can handle: |
|
|
- Simple aggregations: "Get total GMV and orders for last 30 days" |
|
|
- Complex joins: "Get seller performance with KAM assignments" |
|
|
- Time-based analysis: "Show monthly GMV trend by industry" |
|
|
- Advanced logic: "Compare Choice vs Non-Choice GMV in Crossborder" |
|
|
|
|
|
## Limitations |
|
|
|
|
|
- Trained specifically for Daraz schema and business logic |
|
|
- May not generalize to other SQL dialects or schemas |
|
|
- Requires Daraz-specific tables to be available |
|
|
|
|
|
## Training Dataset |
|
|
|
|
|
Custom dataset of 20 SQL query examples covering: |
|
|
- Revenue and GMV analysis |
|
|
- Product performance metrics |
|
|
- Seller segmentation |
|
|
- Category and brand analysis |
|
|
- Time-based trends |
|
|
|
|
|
## Citation |
|
|
|
|
|
If you use this model, please cite: |
|
|
|
|
|
``` |
|
|
@misc{drz-sql-llama3, |
|
|
author = {Bilal326}, |
|
|
title = {drz-sql-llama3: Daraz SQL Generation Model}, |
|
|
year = {2025}, |
|
|
publisher = {HuggingFace}, |
|
|
url = {https://huggingface.co/Bilal326/drz-sql-llama3} |
|
|
} |
|
|
``` |
|
|
|
|
|
## Acknowledgments |
|
|
|
|
|
- Built with [Unsloth](https://github.com/unslothai/unsloth) |
|
|
- Based on Meta's Llama 3 |
|
|
- Fine-tuned for Daraz e-commerce analytics |
|
|
|
