ayeshmaffi/entsql-adapters

EntSQL adapters — cross-dialect and cross-domain text-to-SQL

Ten LoRA adapters for Qwen/Qwen2.5-Coder-1.5B-Instruct from the paper "Cross-dialect and cross-domain generalization of small text-to-SQL models: an execution-verified study with an enterprise benchmark". Code, the 400 EntSQL questions, splits and full experiment logs: https://github.com/ayeshamafrah-am/entsql-dialects.

All adapters: LoRA r=16, α=32, dropout 0.05, all attention and MLP projections, bf16, trained on one laptop GPU (RTX 5070 Ti, 12 GB).

Folder	Training data	Purpose
p2_sft_sqlite	Spider train + BIRD subset, SQLite only	the usual fine-tune (baseline)
p2_sft_multi	same questions, SQLite + MySQL + PostgreSQL, execution-verified	multi-dialect model (main)
p2_sft_multi_notag	as p2_sft_multi, prompt without the dialect tag	ablation
p2_adapt_k30	p2_sft_multi + 30 EntSQL questions per domain	few-shot domain adaptation
p2_adapt_k10	p2_sft_multi + 10 EntSQL questions per domain	few-shot domain adaptation
p2_lodo_<domain> (×5)	p2_sft_multi + EntSQL questions from the 4 other domains	leave-one-domain-out

Results (execution accuracy, %)

Model	Spider dev (SQLite / MySQL / PG)	BIRD dev	EntSQL test (250)
zero-shot	61.0 / 57.8 / 57.6	23.5 / 21.3 / 21.2	20.0 / 23.1 / 21.7
p2_sft_sqlite	74.6 / 65.6 / 67.5	35.5 / 29.9 / 28.3	23.6 / 23.1 / 18.8
p2_sft_multi	76.8 / 76.0 / 75.7	39.2 / 38.8 / 38.7	27.6 / 22.7 / 22.9
p2_adapt_k30	–	–	44.0 / 43.7 / 45.4
p2_lodo_* (mean)	–	–	39.2 / 38.5 / 38.4

Usage

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

base = "Qwen/Qwen2.5-Coder-1.5B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, dtype=torch.bfloat16, device_map="cuda")
model = PeftModel.from_pretrained(model, "ayeshmaffi/entsql-adapters", subfolder="p2_sft_multi")

The prompt format (dialect tag, dialect-specific DDL, sample values, evidence, question) is built by infer_p2.py::build_messages in the GitHub repository; use the same builder for inference.

Citation

Paper submitted to the International Journal of Computing and Digital Systems (2026). Citation will be added on acceptance.