Shizu0n/phi3-mini-sql-generator-merged

Phi-3 Mini SQL Generator — Merged Model

This is the merged (standalone) version for Shizu0n/phi3-mini-sql-generator on top of microsoft/Phi-3-mini-4k-instruct.

The LoRA adapter weights have been merged directly into the base model, making this a standard AutoModelForCausalLM — no PEFT dependency required for inference.

Why two versions?

Repo	Purpose
Shizu0n/phi3-mini-sql-generator	Original QLoRA adapter — documents the training pipeline
Shizu0n/phi3-mini-sql-generator-merged	Merged standalone model — used for deployment and inference

Evaluation — Base vs Fine-tuned

Evaluated on 200 held-out examples from b-mc2/sql-create-context.

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

Training Details

• Dataset: b-mc2/sql-create-context — 1,000 train / 200 validation examples
• Method: QLoRA (4-bit NF4, LoRA rank 16, alpha 32)
• Hardware: NVIDIA T4 (Google Colab free tier)
• Training time: ~21 min
• Final train loss: 0.6526

Validation

The merge was accepted only after all three smoke tests returned a concrete SQL query:

1. PEFT adapter loaded on the base model.
2. Local merged directory after merge_and_unload() and save_pretrained().
3. Downloaded model from this Hugging Face repo with force_download=True.

Reference smoke output:

SELECT AVG(salary), department FROM employees GROUP BY department

Inference example

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

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

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=False)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.float16,
    device_map="auto",
    trust_remote_code=False,
    attn_implementation="eager",
)
model.config.use_cache = False

prompt = (
    "Given the following SQL table, write a SQL query.

"
    "Table: employees (id, name, department, salary)

"
    "Question: What is the average salary per department?

SQL:"
)

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,
    )

text = tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True)
print(text.strip())

Notes

• Use trust_remote_code=False; the built-in Transformers Phi-3 implementation avoids stale remote-code failures.
• Do not patch rope_scaling manually.
• Do not mutate _tied_weights_keys before saving.