README.md

---
library_name: transformers
base_model:
- Qwen/Qwen2.5-Coder-7B-Instruct
pipeline_tag: text-generation
---

## Usage


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


MODEL_NAME = "alpecevit/Qwen2.5-Coder-7B-Instruct-text2sql"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    torch_dtype="auto",
    device_map="auto"
)

model = PeftModel.from_pretrained(model, MODEL_NAME)
model.eval()

test_data = {
    'db_id': 'cre_Students_Information_Systems',
    'schema': 'CREATE TABLE Students (\n`student_id` INTEGER NOT NULL,\n`bio_data` VARCHAR(255) NOT NULL,\n`student_details` VARCHAR(255) NOT NULL,\nPRIMARY KEY (`student_id`)\n)\nCREATE TABLE Transcripts (\n`transcript_id` INTEGER NOT NULL,\n`student_id` INTEGER NOT NULL,\n`date_of_transcript` DATETIME(3),\n`transcript_details` VARCHAR(255) NOT NULL,\nPRIMARY KEY (`transcript_id`),\nFOREIGN KEY (student_id) REFERENCES Students (student_id)\n)\nCREATE TABLE Behaviour_Monitoring (\n`behaviour_monitoring_id` INTEGER NOT NULL,\n`student_id` INTEGER NOT NULL,\n`behaviour_monitoring_details` VARCHAR(255) NOT NULL,\nPRIMARY KEY (`behaviour_monitoring_id`),\nFOREIGN KEY (student_id) REFERENCES Students (student_id)\n)\nCREATE TABLE Addresses (\n`address_id` INTEGER NOT NULL,\n`address_details` VARCHAR(255) NOT NULL,\nPRIMARY KEY (`address_id`)\n)\nCREATE TABLE Ref_Event_Types (\n`event_type_code` CHAR(10) NOT NULL,\n`event_type_description` VARCHAR(255) NOT NULL,\nPRIMARY KEY (`event_type_code`)\n)\nCREATE TABLE Ref_Achievement_Type (\n`achievement_type_code` CHAR(15) NOT NULL,\n`achievement_type_description` VARCHAR(80),\nPRIMARY KEY (`achievement_type_code`)\n)\nCREATE TABLE Ref_Address_Types (\n`address_type_code` CHAR(10) NOT NULL,\n`address_type_description` VARCHAR(255) NOT NULL,\nPRIMARY KEY (`address_type_code`)\n)\nCREATE TABLE Ref_Detention_Type (\n`detention_type_code` CHAR(10) NOT NULL,\n`detention_type_description` VARCHAR(80),\nPRIMARY KEY (`detention_type_code`)\n)\nCREATE TABLE Student_Events (\n`event_id` INTEGER NOT NULL,\n`event_type_code` CHAR(10) NOT NULL,\n`student_id` INTEGER NOT NULL,\n`event_date` DATETIME(3),\n`other_details` VARCHAR(255) NOT NULL,\nPRIMARY KEY (`event_id`),\nFOREIGN KEY (student_id) REFERENCES Students (student_id),\nFOREIGN KEY (event_type_code) REFERENCES Ref_Event_Types (event_type_code)\n)\nCREATE TABLE Teachers (\n`teacher_id` INTEGER NOT NULL,\n`teacher_details` VARCHAR(255),\nPRIMARY KEY (`teacher_id`)\n)\nCREATE TABLE Student_Loans (\n`student_loan_id` INTEGER NOT NULL,\n`student_id` INTEGER NOT NULL,\n`date_of_loan` DATETIME(3),\n`amount_of_loan` DECIMAL(15,4),\n`other_details` VARCHAR(255),\nPRIMARY KEY (`student_loan_id`),\nFOREIGN KEY (student_id) REFERENCES Students (student_id)\n)\nCREATE TABLE Classes (\n`class_id` INTEGER NOT NULL,\n`student_id` INTEGER NOT NULL,\n`teacher_id` INTEGER NOT NULL,\n`class_details` VARCHAR(255) NOT NULL,\nPRIMARY KEY (`class_id`),\nFOREIGN KEY (student_id) REFERENCES Students (student_id),\nFOREIGN KEY (teacher_id) REFERENCES Teachers (teacher_id)\n)\nCREATE TABLE Students_Addresses (\n`student_address_id` INTEGER NOT NULL,\n`address_id` INTEGER NOT NULL,\n`address_type_code` CHAR(10) NOT NULL,\n`student_id` INTEGER NOT NULL,\n`date_from` DATETIME(3),\n`date_to` DATETIME(3),\nPRIMARY KEY (`student_address_id`),\nFOREIGN KEY (student_id) REFERENCES Students (student_id),\nFOREIGN KEY (address_id) REFERENCES Addresses (address_id),\nFOREIGN KEY (address_type_code) REFERENCES Ref_Address_Types (address_type_code)\n)\nCREATE TABLE Detention (\n`detention_id` INTEGER NOT NULL,\n`detention_type_code` CHAR(10) NOT NULL,\n`student_id` INTEGER NOT NULL,\n`datetime_detention_start` DATETIME(3),\n`datetime_detention_end` DATETIME(3),\n`detention_summary` VARCHAR(255),\n`other_details` VARCHAR(255),\nPRIMARY KEY (`detention_id`),\nFOREIGN KEY (student_id) REFERENCES Students (student_id),\nFOREIGN KEY (detention_type_code) REFERENCES Ref_Detention_Type (detention_type_code)\n)\nCREATE TABLE Achievements (\n`achievement_id` INTEGER NOT NULL,\n`achievement_type_code` CHAR(15) NOT NULL,\n`student_id` INTEGER NOT NULL,\n`date_achievement` DATETIME(3),\n`achievement_details` VARCHAR(255),\n`other_details` VARCHAR(255),\nPRIMARY KEY (`achievement_id`),\nFOREIGN KEY (student_id) REFERENCES Students (student_id),\nFOREIGN KEY (achievement_type_code) REFERENCES Ref_Achievement_Type (achievement_type_code)\n)',
    'question': 'What are the type code, details, and date of each achievement?'
}

system_prompt = (
    "You are an SQL query generator that converts natural language inputs into SQL queries, "
    "strictly adhering to the provided database schema. Your task is to:\n\n"
    "Carefully analyze the natural language input to understand its requirements. "
    "Generate a SQL query that retrieves the correct results based only on the database schema provided.\n\n"
    "Rules:\n\n"
    "Do NOT use any tables, columns, or data outside the provided database schema.\n"
    "Ensure the SQL query is syntactically correct and efficient.\n"
    "If the input involves ambiguity or missing details, "
    "assume the most common or straightforward interpretation.\n\n"
    "Respond with only the SQL query, formatted cleanly and consistently. "
    "Do not include explanations, comments, or extra text."
)

user_prompt = (
    "Database schema: {schema}\n\n"
    "Natural language input: {user_query}\n\n"
    "Answer: " 
)

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_prompt.format(schema=test_data['schema'], user_query=test_data['question'])}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512
)

generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
```