license, frameworks, tasks, model-type, domain, language, tags, base_model, base_model_relation, datasets

license	frameworks	tasks	model-type	domain	language	tags	base_model	base_model_relation	datasets
MIT	Pytorch	text-generation	qwen2	nlp	zh, en	LoRA NL2SQL	Qwen/Qwen2.5-Coder-1.5B-Instruct	finetune	ruohuaw/sql-cot

notes

• we recommand turning off beam search and setting temperature = 0.1 for best performance

dataset

• sql-cot: https://modelscope.cn/datasets/ruohuaw/sql-cot

trainer logs

• https://wandb.ai/wangruohua1999econ-none/huggingface/runs/o0nwnyx3?nw=nwuserwangruohua1999econ
• https://swanlab.cn/@ruohua/deepquery-v4-sft/runs/cqgymo8f4asr76i0gpol6/chart

usage:

import torch
from modelscope import snapshot_download
from transformers import AutoTokenizer, AutoModelForCausalLM
model_dir = snapshot_download('ruohuaw/deepquery-1.5b-sft')

# 系统提示设置
sys_prompt = """You are DeepQuery, a data science expert. Below, you are presented with a database schema, a question and a hint.Your task is to read the schema with annotations of the columns, understand the question and the hint, and generate a valid SQL query to answer the question. You should reason step by step, and includes your reasonings between <think> and </think>."""

tokenizer = AutoTokenizer.from_pretrained(model_dir)
model = AutoModelForCausalLM.from_pretrained(model_dir)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)
model.eval()
def generate_response(input_prompt, temperature = 0.1, max_length=512*4, model = model, tokenizer = tokenizer):
    inputs = tokenizer(input_prompt, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_length = max_length,
            temperature = temperature
        )
    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return response

query = """Database Schema
###
CREATE TABLE Country
(
	CountryCode TEXT not null primary key,
	LongName TEXT, -- `Long Name` description: long or full name of countries
);

CREATE TABLE Series
(
	SeriesCode TEXT not null primary key,
);

CREATE TABLE SeriesNotes
(
	Seriescode TEXT not null, -- `Series code` description: code identifying the series
	Year TEXT not null, --
	Description TEXT, --
	primary key (Seriescode, Year),
	foreign key (Seriescode) references Series(SeriesCode),
);

CREATE TABLE CountryNotes
(
	Countrycode TEXT NOT NULL, --
	Seriescode TEXT NOT NULL, -- `Series code` description: Series code of countries
	Description TEXT, --
	primary key (Countrycode, Seriescode),
	FOREIGN KEY (Seriescode) REFERENCES Series(SeriesCode),
	FOREIGN KEY (Countrycode) REFERENCES Country(CountryCode),
);

###
Question: 
Please list the full names of any three countries that have their series code with a description of UN Energy Statistics (2014). 

Hint:
full name refers to longname
"""
input_prompt = tokenizer.apply_chat_template(
                [
                    {"role": "system", "content": SYS},
                    {"role": "user", "content": QUERY},
                    
                ],
                tokenize=False,
                add_generation_prompt=True
            )
TRUE_ANSWER="""SELECT DISTINCT T2.LongName FROM CountryNotes AS T1 INNER JOIN Country AS T2 ON T1.Countrycode = T2.CountryCode WHERE T1.Description = 'Sources: UN Energy Statistics (2014)' LIMIT 3"""

RESPONSE = generate_response(input_prompt)
print(f"Response: \n{RESPONSE}\nTrue answer: \n{TRUE_ANSWER}")

SDK下载

#安装ModelScope
pip install modelscope

#SDK模型下载
from modelscope import snapshot_download
model_dir = snapshot_download('ruohuaw/deepquery-1.5b-sft')

Git下载

#Git模型下载
git clone https://www.modelscope.cn/ruohuaw/deepquery-1.5b-sft.git