metadata
language:
- en
tags:
- text-to-sql
- phi-3
- genbi
- dbt
- azure-sql
- qlora
base_model: microsoft/Phi-3-mini-4k-instruct
license: mit
GenBI Phi-3 SQL Agent
Model Description
This model is a specialized Small Language Model (SLM) designed to act as the reasoning engine for an Agentic GenBI Enterprise Platform. It translates natural language business questions into highly accurate, executable T-SQL queries.
It has been instruction-tuned to understand modern data stack semantics, specifically bridging the gap between a dbt Semantic Layer and an Azure SQL data warehouse.
- Developed by: deepinfo
- Model type: Causal Language Model (Fine-tuned via QLoRA)
- Base model: Microsoft Phi-3-mini-4k-instruct
- Primary Use Case: Enterprise Business Intelligence, Autonomous SQL Generation, and LangGraph Agentic Workflows.
Training Details
This model was fine-tuned using a synthetic dataset generated from a strictly typed dbt semantic layer (manifest.json). The training ensures the model adheres strictly to predefined business logic (e.g., Margin, Revenue, Customer Lifetime Value) rather than hallucinating column names or relationships.
- Hardware: NVIDIA T4/A100 (Google Colab)
- Technique: QLoRA (Quantized Low-Rank Adaptation)
- Format: Merged fp16/bf16 weights.
Usage Example (Python)
from transformers import pipeline
pipe = pipeline("text-generation", model="deepinfo/genbi-phi3-sql-agent")
prompt = """You are the reasoning engine for an Agentic GenBI Enterprise Platform. Your role is to translate business questions into accurate T-SQL queries for an Azure SQL database.
Context:
Target Table: fct_sales_performance
Columns:
- gross_revenue: Standard definition for total sales: Qty * Price.
- gross_profit: Margin definition: Revenue - Cost.
- category_name: The top-level classification of the product sold.
User: What was the total gross profit generated by the Bikes category?"""
output = pipe(prompt, max_new_tokens=100, temperature=0.1)
print(output[0]['generated_text'])