---
base_model: microsoft/Phi-3-mini-4k-instruct
library_name: transformers
license: mit
language:
- en
datasets:
- b-mc2/sql-create-context
tags:
- sql
- text-to-sql
- code-generation
- phi-3
- fine-tuned
- text-generation
- phi3
pipeline_tag: text-generation
---

# Phi-3 Mini SQL Generator — Merged Model

Merged standalone version of [Shizu0n/phi3-mini-sql-generator](https://huggingface.co/Shizu0n/phi3-mini-sql-generator)
— LoRA adapter weights fused into [Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct).
No PEFT dependency required for inference.

## 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 model (fine-tuned)** | **73.5%** |

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

## Why two versions?

| Repo | Purpose |
|---|---|
| [`Shizu0n/phi3-mini-sql-generator`](https://huggingface.co/Shizu0n/phi3-mini-sql-generator) | QLoRA adapter — documents the training pipeline |
| `Shizu0n/phi3-mini-sql-generator-merged` | Merged standalone — used for deployment and inference |

## Training Details

- **Dataset:** b-mc2/sql-create-context — 1,000 train / 200 validation examples
- **Method:** QLoRA (4-bit NF4, LoRA rank 16, alpha 32, target modules: qkv_proj/o_proj/gate_up_proj/down_proj)
- **Hardware:** NVIDIA T4 (Google Colab free tier)
- **Training time:** ~21 min
- **Final train loss:** 0.6526
- **Best checkpoint:** step 250 (by eval loss)

## Inference Example

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

model_id = "Shizu0n/phi3-mini-sql-generator-merged"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=False,
    attn_implementation="eager",
)
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=80,
        do_sample=False,
        use_cache=False,
        repetition_penalty=1.1,
        pad_token_id=tokenizer.eos_token_id,
    )
prompt_len = inputs["input_ids"].shape[-1]
print(tokenizer.decode(outputs[0][prompt_len:], skip_special_tokens=True))
```

Expected output:
```sql
SELECT AVG(salary), department FROM employees GROUP BY department
```

## Validation

Merge accepted after three smoke tests:
1. PEFT adapter loaded on base model
2. Local merged directory after `merge_and_unload()` + `save_pretrained()`
3. Downloaded from this repo with `force_download=True`

## 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