llama-3.1-8b-atlassql-br

LoRA adapter of meta-llama/Llama-3.1-8B-Instruct fine-tuned on AtlasSQL-BR, a Brazilian Portuguese geospatial Text-to-SQL dataset over the CulturaEduca PostGIS database (schools, public facilities and the seven-level IBGE territorial hierarchy). The model translates a question in Portuguese into a PostGIS SQL query. No schema is injected in the prompt: the adapter is specific to the CulturaEduca schema.

This is the model reported in the master's thesis AtlasSQL-BR: a Brazilian Portuguese geospatial Text-to-SQL dataset (IME-USP, 2026); a first version of the experiment was published at SBBD 2026.

Training

Setting	Value
Data	thesis-split tag of the dataset: 784 training / 196 validation pairs, stratified by complexity tier, spatial function and territorial division
Prompt	Traduza para SQL: {question} (no schema)
LoRA	q_proj, v_proj; r = 8, alpha = 32, dropout = 0.1 (< 0.5 % trainable parameters)
Optimization	10 epochs, batch 2 x 2 accumulation, AdamW, lr 5e-5 (linear), weight decay 0.01, 612 warmup steps, bf16, gradient checkpointing
Checkpoint	epoch 6, minimum validation loss (0.172; base model 1.180)
Hardware	Apple MacBook Pro, M5 Pro, 48 GB unified memory (PyTorch MPS); 2 h 44 min

Results (196-pair validation split)

Metric	Base model	This adapter
Execution Accuracy (%)	0.0	15.8
Executable Rate (%)	0.0	52.6
Token F1	0.285	0.757
Structural F1	0.169	0.753
Geospatial Function F1	0.108	0.639
Spatial Exact Match (%)	0.5	28.1

Full breakdowns (by tier, territorial division and spatial function) and the evaluation code are in the code repository.

Usage

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

base = "meta-llama/Llama-3.1-8B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(model, "datafromlopes/llama-3.1-8b-atlassql-br").eval()

prompt = "Traduza para SQL: Quais escolas estão a até 2 km da biblioteca municipal de Campinas?"
ids = tok(prompt, return_tensors="pt").to(model.device)
out = model.generate(**ids, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))

The output should be post-processed as in the evaluation code (strip code fences and annotation tags, keep the first WITH/SELECT statement).

Citation

@inproceedings{lopes2026atlassql,
  title     = {AtlasSQL-BR: A Brazilian Portuguese Geospatial Text-to-SQL Dataset with Spatial Hierarchies},
  author    = {Lopes, Diego O. and Braghetto, Kelly R.},
  booktitle = {Proceedings of the 41st Brazilian Symposium on Databases (SBBD)},
  year      = {2026}
}

License: MIT for the adapter weights; the base model is subject to the Llama 3.1 Community License.