BART-based transformer that translates natural-language questions into executable SQL over user-uploaded CSV files.
Business users can't write SQL but need to query their data. Off-the-shelf LLMs hallucinate column names and produce non-executable queries on unseen schemas.
Fine-tuned facebook/bart-base on the Spider cross-domain text-to-SQL benchmark with schema-aware prompting and entity resolution. Inference optimized via ONNX export.
| Metric | Value |
|---|---|
| Spider exact-match accuracy | 45.6% |
| Inference latency (ONNX vs PyTorch) | 60% faster |
| Live demo response time | < 200 ms |
- User uploads one or more CSV files → loaded into in-memory SQLite.
- Schema is serialized and prepended to the user question.
- Fine-tuned BART generates the SQL query.
- Post-processing: regex correction, fuzzy matching for column names.
- Query executes against SQLite; result returned as a table.
- Model: BART-base, fine-tuned on Spider
- Serving: FastAPI + Gradio on Hugging Face Spaces
- Optimization: ONNX runtime (60% speedup vs PyTorch baseline)
- Data: Spider benchmark + user-uploaded CSVs
git clone https://github.com/Rushikesh-S-Ware/nlp-sql-transformer
cd nlp-sql-transformer
pip install -r requirements.txt
python src/app.pyOpen http://localhost:7860 and upload a CSV.
👉 https://huggingface.co/spaces/Rushikesh-S-Ware/NLP-SQL-Transformer
| Path | Purpose |
|---|---|
src/ |
FastAPI server + inference pipeline |
tests/ |
Unit and integration tests |
.github/workflows/ |
CI: lint, test, build |
Model_Training.ipynb |
Fine-tuning pipeline on Spider |
Demo_Notebook.ipynb |
Usage examples |
Model_Checkpoint/ |
Fine-tuned weights |
requirements.txt |
Dependencies |
.env.example |
Required environment variables |
MIT