curie-backend 0.1.0

Text2SQL Assistant Backend

Project Curie - Text2SQL

Project description

• Idea: Use natural language prompts to generate SQL queries

• Front-end: VS Code Plugin

  • Integration of trained LLM into useful plugin with quick insights
  • Establish connection with existing database
  • Use natural language to describe queries to this database, which are generated by fine-tuned LLM
  • Ability to run generated SQL code directly via plugin to see results immediately with no bothersome copying

• ...

Simple example:

Project structure

evaluation/   Benchmark setups for Text-to-SQL evaluation
data/         Local datasets and sample data
docs/         Project documentation, presentations, and research notes
experiments/  Ongoing fine-tuning experiments and run notes
extension/    VS Code extension and demo database
ml/           Training, fine-tuning, and inference code
ml/models/    Model metadata, configs, and model cards

Project team and responsibilities

Name	E-Mail	Identifier	Responsibilities / tasks
Tamer Lewis	[e-mail]	go68maj	TBD
Antonia Gilles	[e-mail]	go34cih	TBD
Leon Pfänder	[e-mail]	ge27fob	TBD
Christoph Klein	[e-mail]	ge42low	TBD

Project setup

The project uses Python 3.12 and a pinned ML stack from requirements.txt.

python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

Verify that the installed dependencies are consistent:

python -m pip check

Git LFS for model weights

Git LFS stores large model binaries on the GitLab LFS server while Git tracks small pointer files. This repository uses LFS for *.safetensors files.

Minimal setup:

brew install git-lfs
git lfs install
git lfs pull

Add and push model weights through the normal Git workflow:

git add .gitattributes ml/models/<model>/adapter_model.safetensors
git commit -m "Add model adapter"
git push

After cloning, Git LFS normally downloads tracked files automatically. If a model file is only an LFS pointer or is missing, run git lfs pull. Verify the available model files with git lfs ls-files.

LFS does not change how the benchmark is invoked:

python3 evaluation/spider/run.py \
  --split dev \
  --model-dir ml/models/<model>

Spider

Download the Spider dataset and place it in the data/ folder:

https://drive.google.com/file/d/1403EGqzIDoHMdQF4c9Bkyl7dZLZ5Wt6J/view

Sources

TBD