Querylingo

Querylingo is a working in process LLM to parse Natural Language Queries to SQL and the other way around. This project is part of my Bachelor Thesis at the University.

Main Goal

The main goal of this project is to create a tool that can parse natural language queries to SQL and the other way around. This tool should be able to understand the user's query and translate it to a SQL query that can be executed in a database.

Getting Started

• Create datasets:

python src/data.py -h

usage: data.py [-h] [--data-dir DATA_DIR] [--download] [--process] [--save {csv,jsonl}] [--export EXPORT EXPORT] [--create-dataset CREATE_DATASET CREATE_DATASET]

Download datasets

options:
  -h, --help            show this help message and exit
  --data-dir DATA_DIR, -d DATA_DIR
                        Directory to save the datasets
  --download, -dw       Download datasets from huggingface
  --process, -p         Process datasets and save them to CSV or JSONL files depending on the flag --save (default: JSONL)
  --save {csv,jsonl}, -s {csv,jsonl}
                        Save processed datasets to CSV or JSONL files
  --export EXPORT EXPORT, -e EXPORT EXPORT
                        Export processed datasets to SQL
  --create-dataset CREATE_DATASET CREATE_DATASET, -cd CREATE_DATASET CREATE_DATASET
                        Path where to find the SQLite DB and path to create the datasets from the processed datasets to JSON or CSV file depending on the flag --save (default: JSONL)

# Example
WORKERS=8 DEBUG=3 python src/data.py -dw -d /mnt/d/tfg/ -e /mnt/d/tfg/processed/hf/ /mnt/d/tfg/processed/ -p -s jsonl -cd /mnt/d/tfg/processed/datasets.sqlite /mnt/d/tfg/processed/datasets/

Dataset can be downloaded from here (hugginface)

• Train model:

python src/train.py -h

usage: train.py [-h] [--model MODEL] [--use_kbit] [--epochs EPOCHS] [--batch_size BATCH_SIZE] [--accumulation_steps ACCUMULATION_STEPS] [--eval_strategy EVAL_STRATEGY]
                [--warmup_steps WARMUP_STEPS] [--eval_steps EVAL_STEPS] [--learning_rate LEARNING_RATE] [--fp16] [--logging_steps LOGGING_STEPS] [--output_dir OUTPUT_DIR]
                [--optim OPTIM] [--text_field TEXT_FIELD] [--max_length MAX_LENGTH]

Train a model

options:
  -h, --help            show this help message and exit
  --model MODEL         Model to train
  --use_kbit            Use kbit training
  --epochs EPOCHS       Number of epochs
  --batch_size BATCH_SIZE
                        Batch size
  --accumulation_steps ACCUMULATION_STEPS
                        Gradient accumulation steps
  --eval_strategy EVAL_STRATEGY
                        Evaluation strategy
  --warmup_steps WARMUP_STEPS
                        Warmup steps
  --eval_steps EVAL_STEPS
                        Evaluation steps
  --learning_rate LEARNING_RATE
                        Learning rate
  --fp16                Use fp16
  --logging_steps LOGGING_STEPS
                        Logging steps
  --output_dir OUTPUT_DIR
                        Output directory
  --optim OPTIM         Optimizer
  --text_field TEXT_FIELD
                        Text field
  --max_length MAX_LENGTH
                        Max length

Create animations

manim -pqh docs/animations/llm.py GeneratingText

License

This project is licensed under the MIT License - see the LICENSE file for details