CodeLlama-7b Text-to-SQL (QLoRA Fine-tune)
A CodeLlama-7b model fine-tuned on the Spider dataset using QLoRA (4-bit quantization + LoRA adapters) for the task of converting natural language questions into SQL queries.
Model Details
Model Description
This model is a parameter-efficient fine-tune of CodeLlama-7b using QLoRA on the Spider Text-to-SQL benchmark dataset. It takes a database schema (as DDL statements) and a natural language question as input, and generates the corresponding SQL query.
- Developed by: [Yash Shukla]
- Model type: Causal Language Model (CodeLlama-7b + LoRA adapters)
- Language(s) (NLP): English
- License: Llama 2 Community License
- Finetuned from model: codellama/CodeLlama-7b-hf
Model Sources
Uses
Direct Use
This model is intended for converting natural language questions into SQL queries given a database schema. It is suitable for:
- Building natural language database interfaces
- SQL query auto-completion tools
- Educational tools for learning SQL
Downstream Use
The LoRA adapter can be merged into the base model or used directly with PEFT for further fine-tuning on domain-specific SQL dialects or private schemas.
Out-of-Scope Use
- Not suitable for production databases without human review of generated queries
- Not tested on non-English questions
- Not designed for NoSQL or non-relational query languages
Bias, Risks, and Limitations
- The model is trained only on Spider which covers ~200 databases — it may struggle with unseen schema patterns
- Generated SQL is not guaranteed to be syntactically or semantically correct
- The model may hallucinate column or table names not present in the provided schema
- All column types are inferred as TEXT/INTEGER/REAL — nuanced type handling may be incorrect
Recommendations
Always validate generated SQL against the actual database before execution. Do not run generated queries directly on production systems without review.
How to Get Started with the Model
1from transformers import AutoTokenizer, AutoModelForCausalLM
2from peft import PeftModel
3import torch
4
5BASE_MODEL = "codellama/CodeLlama-7b-hf"
6ADAPTER = "your-username/codellama-7b-text2sql" # this repo
7
8tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
9base = AutoModelForCausalLM.from_pretrained(BASE_MODEL, torch_dtype=torch.float16, device_map="auto")
10model = PeftModel.from_pretrained(base, ADAPTER)
11
12schema = "CREATE TABLE employees (id INTEGER, name TEXT, salary REAL, dept_id INTEGER);\nCREATE TABLE departments (id INTEGER, dept_name TEXT);"
13question = "List all employees in the Engineering department ordered by salary descending."
14
15prompt = (
16 f"<s>[INST] You are an expert SQL assistant. "
17 f"Given the database schema below, write a SQL query that answers the question.\n\n"
18 f"### Schema:\n{schema}\n\n"
19 f"### Question:\n{question} [/INST]\n"
20)
21
22inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
23outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.1, do_sample=False)
24sql = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
25print(sql)
Training Details
Training Data
- Dataset: Spider — a large-scale human-labeled Text-to-SQL dataset
- Train split: ~7,000 examples
- Validation split: ~1,034 examples
- Filtering: Examples exceeding 1024 tokens after prompt formatting were excluded
Training Procedure
Preprocessing
Each example was formatted using an Alpaca-style instruction template with the database DDL schema and natural language question as input, and the SQL query as output. SQL queries were normalized by collapsing whitespace and stripping trailing semicolons.
Training Hyperparameters
- Training regime: fp16 mixed precision
- Quantization: 4-bit NF4 (bitsandbytes)
- LoRA rank (r): 16
- LoRA alpha: 32
- LoRA dropout: 0.05
- Target modules:
q_proj, v_proj
- Epochs: 3
- Batch size: 4 per device + 4 gradient accumulation steps (effective batch = 16)
- Learning rate: 2e-4 with cosine schedule
- Warmup ratio: 0.03
- Optimizer: paged_adamw_32bit
- Max sequence length: 1024
Speeds, Sizes, Times
- Hardware: Kaggle T4 x2 GPU
- Training time: ~2.5 hours for 3 epochs
- Adapter size: ~84 MB
Evaluation
Testing Data, Factors & Metrics
Testing Data
Spider validation split (~1,034 examples across unseen databases).
Metrics
- Eval loss monitored during training via
evaluation_strategy="steps" every 200 steps
- Best checkpoint selected by lowest eval loss
Results
| Metric | Value |
|---|
| Eval Loss | [0.160100] |
Summary
The model learns to ground SQL generation in the provided schema, producing syntactically valid queries for common SQL patterns including SELECT, WHERE, JOIN, GROUP BY, and ORDER BY.
Environmental Impact
- Hardware Type: NVIDIA Tesla T4 x2
- Hours used: ~2.5
- Cloud Provider: Kaggle (Google Cloud)
- Compute Region: us-central1
- Carbon Emitted: ~0.2 kg CO2eq (estimated via ML Impact Calculator)
Technical Specifications
Model Architecture and Objective
- Base: CodeLlama-7b (decoder-only transformer, 7B parameters)
- Adapter: LoRA applied to
q_proj and v_proj attention layers
- Objective: Causal language modeling (next token prediction) on formatted SQL instruction examples
Compute Infrastructure
Hardware
2x NVIDIA Tesla T4 (16 GB VRAM each) on Kaggle free tier
Software
- Python 3.12
- PyTorch 2.2.0
- Transformers 4.40.0
- PEFT 0.10.0
- TRL 0.9.6
- bitsandbytes 0.43.1
- Accelerate 0.29.3
Citation
BibTeX:
1@inproceedings{yu2018spider,
2 title = {Spider: A Large-Scale Human-Labeled Dataset for Complex and Cross-Domain Semantic Parsing and Text-to-SQL Task},
3 author = {Yu, Tao and Zhang, Rui and Yang, Kai and Yasunaga, Michihiro and Wang, Dongxu and Li, Zifan and Ma, James and Li, Irene and Yao, Qingning and Roman, Shanelle and others},
4 booktitle = {EMNLP},
5 year = {2018}
6}
Model Card Authors
[Yash Shukla]
Model Card Contact
