sharadsin/Mixtral-8x7B-sql-ft-v1

SQL-Converter Mixtral 8x7B v0.1

Convert Natural Language to SQL

Overview

Mixtral-8x7B-sql-ft-v1 is fine-tuned from Mixtral 8x7B to convert natural language to SQL queries.

Base Model

mistralai/Mixtral-8x7B-v0.1

Fine-Tuning

• Dataset: 5,000 natural language-SQL pairs.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

base_model_id = 'mistralai/Mixtral-8x7B-v0.1'
adapter_id = 'sharadsin/Mixtral-8x7B-sql-ft-v1'

bnb_config = BitsAndBytesConfig(
  load_in_4bit = True,
  bnb_4bit_use_double_quant = True,
  bnb_4bit_compute_dtype = torch.bfloat16,
  bnb_4bit_quant_type = "nf4",
)

base_model = AutoModelForCausalLM.from_pretrained(
  base_model_id,
  quantization_config = bnb_config,
  device_map = "auto",
  trust_remote_code = True,
)
tokenizer = AutoTokenizer.from_pretrained(base_model_id, add_bos_token = True, trust_remote_code = True)

ft_model = PeftModel.from_pretrained(base_model, adapter_id)

eval_prompt= """SYSTEM: Use the following contextual information to concisely answer the question.

USER: CREATE TABLE EmployeeInfo (EmpID INTEGER, EmpFname VARCHAR, EmpLname VARCHAR, Department VARCHAR, Project VARCHAR,Address VARCHAR, DOB DATE, Gender CHAR)
===
Write a query to fetch details of employees whose EmpLname ends with an alphabet 'A' and contains five alphabets?
ASSISTANT:"""

model_input = tokenizer(eval_prompt, return_tensors="pt").to("cuda")

ft_model.eval()
with torch.inference_mode():
    print(tokenizer.decode(ft_model.generate(**model_input, max_new_tokens=70,top_k=4, penalty_alpha = 0.6, repetition_penalty=1.15)[0], skip_special_tokens= False))

Limitations

• Less accurate with very complex queries.
• Generates extra gibberish content after providing the answers.

Framework versions

• PEFT 0.7.1