README.md

---
base_model: microsoft/Phi-3-mini-4k-instruct
library_name: peft
license: mit
language:
- en
datasets:
- b-mc2/sql-create-context
tags:
- sql
- code-generation
- text-to-sql
- phi-3
- lora
- qlora
- fine-tuned
- peft
pipeline_tag: text-generation
---

# Phi-3 Mini SQL Generator (QLoRA Fine-tuned)

Fine-tuned version of [Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct)
for **natural language → SQL** generation using QLoRA on a T4 GPU (Google Colab, ~20 min).

## Evaluation — Base vs Fine-tuned

Evaluated on 200 held-out examples from [b-mc2/sql-create-context](https://huggingface.co/datasets/b-mc2/sql-create-context).

| Model | Exact Match |
|---|---|
| Phi-3-mini-4k-instruct (base) | 2.0% |
| **This adapter (fine-tuned)** | **73.5%** |

> Exact match: normalized SQL comparison (lowercase, strip whitespace/semicolons).

## Training Details

- **Dataset:** b-mc2/sql-create-context — 1,000 train / 200 validation examples
- **Epochs:** 3
- **Effective batch size:** 8
- **Learning rate:** 0.0002
- **Max sequence length:** 512
- **Hardware:** NVIDIA T4 (Google Colab free tier)
- **Training time:** 21.2 min
- **Final train loss:** 0.6526
- **Best checkpoint:** step 250 (lowest eval loss — mild overfitting observed after epoch 2)

## LoRA Config

| Parameter | Value |
|---|---|
| Rank (r) | 16 |
| Alpha | 32 |
| Dropout | 0.05 |
| Target modules | `qkv_proj`, `o_proj`, `gate_up_proj`, `down_proj` |
| Quantization | 4-bit NF4 (QLoRA) |

## How to Use

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

tokenizer = AutoTokenizer.from_pretrained(
    "microsoft/Phi-3-mini-4k-instruct", trust_remote_code=True
)
base_model = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-3-mini-4k-instruct",
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
    attn_implementation="eager",
)
model = PeftModel.from_pretrained(base_model, "Shizu0n/phi3-mini-sql-generator")
model.eval()

prompt = (
    "Given the following SQL table, write a SQL query.\n\n"
    "Table: employees (id, name, department, salary)\n\n"
    "Question: What is the average salary per department?\n\nSQL:"
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.inference_mode():
    outputs = model.generate(**inputs, max_new_tokens=100, do_sample=False)
prompt_len = inputs["input_ids"].shape[-1]
print(tokenizer.decode(outputs[0][prompt_len:], skip_special_tokens=True))
```

## Related

The LoRA adapter weights have been merged into a standalone model at
[Shizu0n/phi3-mini-sql-generator-merged](https://huggingface.co/Shizu0n/phi3-mini-sql-generator-merged)
— no PEFT dependency required for inference.

## Limitations

- Fine-tuned on 1,000 examples — best suited for simple to medium complexity SELECT queries
- Not tested on dialect-specific SQL (PostgreSQL/MySQL-specific functions)
- May struggle with multi-table JOINs and nested subqueries