This repository contains the final LoRA adapter for the Qwen 3.5 2B model line used in a master's thesis project on local large language models for NL2SQL generation.

The adapter generates SQLite queries from natural-language questions and a relational database schema. It must be loaded together with the corresponding base model.

Base model

Model: Qwen/Qwen3.5-2B-Base
Revision: b1485b2fa6dfa1287294f269f5fb618e03d52d7c

Adapter

Method: LoRA supervised fine-tuning
LoRA rank: 8
LoRA alpha: 16
LoRA dropout: 0.05
Target modules: all suitable linear modules
Quantization during training: none
Maximum training sequence length: 2,048 tokens
Best checkpoint: checkpoint-502
The published root adapter corresponds to the selected best checkpoint.

SHA-256 of adapter_model.safetensors:

6b92f120365d127d0c51a4c532953207d65cff611ac08cb7d573880be18223f3

Training configuration

Training examples: 25,000
Spider Train examples: 6,960
SQL Create Context examples: 18,040
Validation set: MixedVal2500-v2
Validation examples: 2,500
Learning rate: 1e-4
Scheduler: constant
Train batch size: 2
Gradient accumulation steps: 4
Effective batch size: 8
Seed: 42
Maximum epochs: 5
Early stopping patience: 2
Early stopping threshold: 0.001
Precision: FP16
Gradient checkpointing: enabled
Attention implementation: FlashAttention 2

Spider Dev was not used for training, validation, early stopping, or checkpoint selection.

Evaluation

The final adapter was evaluated on all 1,032 Spider Dev cases.

Zero-shot evaluation:

Execution Match Accuracy: 59.59% (615/1,032)
Execution Success Rate: 85.37% (881/1,032)
Maximum input length: 2,048 tokens
Maximum generated tokens: 256

These values correspond to the authoritative evaluation run documented in the accompanying GitHub repository.

Intended prompt behavior

The model is instructed to return only a valid SQLite query:

no explanation
no Markdown
no comments
no unnecessary tables or columns
only SELECT or WITH queries
output terminated with a semicolon

The exact prompt construction and evaluation pipeline are documented in the project repository.

Loading

python

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

base_model_id = "Qwen/Qwen3.5-2B-Base"
adapter_id = "mehmet1899/qwen35-2b-nl2sql-lora"
adapter_revision = "c6373ca847220b446d3d84859e914f89dc208375"

tokenizer = AutoTokenizer.from_pretrained(
    base_model_id,
    revision="b1485b2fa6dfa1287294f269f5fb618e03d52d7c",
)

model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    revision="b1485b2fa6dfa1287294f269f5fb618e03d52d7c",
    torch_dtype=torch.float16,
    device_map="auto",
)

model = PeftModel.from_pretrained(
    model,
    adapter_id,
    revision=adapter_revision,
)
model.eval()

Reproducibility

Code, training and evaluation configurations, environment information, run manifests, and result summaries are available at:

https://github.com/md181099/nl2sql-masterthesis

The files training_metadata.json, training_history.csv, and training_history.jsonl provide additional training provenance.

Limitations

The adapter was evaluated primarily on the Spider benchmark and SQLite databases.
Performance on other database systems or unseen schema conventions is not guaranteed.
Execution Match depends on the database contents and the execution-based evaluation procedure.
The model may still generate invalid, incomplete, or semantically incorrect SQL.
The adapter should not be used to execute unrestricted queries against production databases without validation and access controls.