---
license: mit
base_model: microsoft/Phi-3-mini-4k-instruct
library_name: peft
pipeline_tag: text-generation
language:
  - en
tags:
  - text-to-sql
  - sql
  - lora
  - qlora
  - peft
  - phi-3
---

# Phi-3-mini Text-to-SQL — LoRA Adapter

A **QLoRA** adapter that specializes [`microsoft/Phi-3-mini-4k-instruct`](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct) (3.8B) for **natural-language → SQLite** generation over a fixed enterprise schema (departments / employees / products / sales).

- 🔌 **9 MB adapter** (0.117% the size of the base model)
- ⚡ Trained in **~3 minutes** within **5.2 GB** of GPU memory on a 6 GB laptop GPU (RTX 4050)
- 🧪 **75% execution-match / 100% valid-SQL** on held-out questions (up from **41.7%** for the base model)
- 📦 Quantized GGUFs for CPU serving: [`Bhuvandesai/phi3-text-to-sql-gguf`](https://huggingface.co/Bhuvandesai/phi3-text-to-sql-gguf)
- 🖥️ Live demo: [`Bhuvandesai/phi3-text-to-sql-studio`](https://huggingface.co/spaces/Bhuvandesai/phi3-text-to-sql-studio)

## How to use

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "microsoft/Phi-3-mini-4k-instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, "Bhuvandesai/phi3-text-to-sql-adapter")

SCHEMA = """You are a Text-to-SQL generator. Given a database schema and a natural language
question, write a valid SQLite query. Output only the raw SQL.

Database Schema:
Table departments(id, name, manager_id)
Table employees(id, name, department_id, salary, hire_date, manager_id)
Table products(id, name, category, price)
Table sales(id, employee_id, product_id, amount, quantity, sale_date)"""

msgs = [{"role": "user", "content": f"{SCHEMA}\n\nQuestion: What is the average salary by department?"}]
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=128, do_sample=False)
print(tok.decode(out[0], skip_special_tokens=True))
```

For CPU / no-GPU use, prefer the quantized GGUFs with `llama.cpp` (see the GGUF repo).

## Training

| | |
|---|---|
| Method | QLoRA (4-bit NF4 + double-quant, bf16 compute) |
| LoRA | r=8, α=16, dropout=0.05, bias=none |
| Trainable params | **4,456,448 (0.1165%** of 3.82B) |
| Data | 50 train / 12 held-out NL→SQL pairs (synthetic schema) |
| Schedule | 3 epochs, effective batch 4, lr 2e-4 cosine, paged_adamw_8bit |
| Hardware | NVIDIA RTX 4050 Laptop (6 GB) |
| Runtime / peak VRAM | 193.7 s / 5.21 GB reserved |

### Results (held-out, greedy decoding)

| Metric | Base Phi-3-mini | **This adapter** |
|---|---:|---:|
| Execution-match (run SQL, compare rows) | 41.7% | **75.0%** |
| Valid SQL rate | 100% | **100%** |
| Eval loss (end of training) | — | **0.0597** (−89.9%) |
| Eval token accuracy | — | **98.4%** |

Strict execution-match is conservative: 2 of the 3 held-out "misses" are reasonable answers with a different column projection than the reference; counting "query correctly answers the question" ≈ **92%**.

## Limitations & honest notes

- **Single fixed schema.** Trained on one synthetic database; it is not a general cross-schema text-to-SQL model.
- **Small dataset (50/12).** Metrics are directional, not statistically tight.
- **LoRA module coverage.** Because Phi-3 fuses `q/k/v` (`qkv_proj`) and gate/up (`gate_up_proj`), PEFT name-matching adapted only `o_proj` and `down_proj` (2 of the 7 listed modules). It still trained well; a future version should target `qkv_proj`/`gate_up_proj` for fuller coverage.

A full write-up (fine-tuning + quantization deep dive with all benchmarks) accompanies this model.

License: MIT (inherits from the base model).