ImobAI Portuguese Intent Baseline

This educational baseline classifies short synthetic Portuguese real-estate messages into eight intents. It combines word TF-IDF, character char_wb TF-IDF, and Logistic Regression.

Version: 1.0.0
Model SHA-256: e657062aa88ae82c998d20f888cafdfc54f073941417d0ae8a3c09454f434743
Dataset content SHA-256: 0313a8d0c850fa18f7c720efca24f178481d2809597c64c64a13d4a6ec9d14ec
Source project: https://github.com/Juliarezende521/Imob-IA
License: Apache License 2.0 (apache-2.0)

License status: LICENSES_DECLARED

Apache 2.0 applies to this model artifact and its package files. It does not replace or alter the licenses of scikit-learn, NumPy, skops, or other libraries.

Architecture and training

word TF-IDF: unigrams and bigrams, Unicode accent stripping, sublinear TF;
character TF-IDF: char_wb, 3–5 characters, min_df=2;
FeatureUnion of both representations;
Logistic Regression: C=1.0, solver="lbfgs", max_iter=2000, seed 42;
final package fit: 256 train + 64 validation records (320 total);
test and final challenge records are excluded from fitting.

The configuration was selected earlier using train and validation only. This package reproduces the baseline initialized by the local API; it does not perform new model selection.

Labels

BUSCA_IMOVEL
AGENDAR_VISITA
DUVIDA_IMOVEL
DOCUMENTACAO_LOCACAO
FINANCEIRO_LOCACAO
MANUTENCAO_IMOVEL
ATENDIMENTO_HUMANO
FORA_DE_ESCOPO

Frozen evaluation results

Evaluation	Accuracy	Macro F1	Correct
Validation during selection	0.766	0.764	—
Previously observed test	0.781	0.786	50/64
Final challenge, opened once	0.682	0.673	131/192

These are previously frozen project results. The challenge was not executed again, is not included in either package, and must not be published. The dataset is small and synthetic, so these values do not imply production readiness or statistical generalization.

The experimental multilingual Transformer performed worse than this baseline on the observed test and challenge evaluations. Its weights are not included. In this project, higher complexity did not produce better observed quality.

Local secure loading

Only load a reviewed local artifact whose SHA-256 matches the manifest. Do not automatically trust types reported by an unknown file.

from pathlib import Path
from skops.io import get_untrusted_types, load

path = Path("model.skops")
untrusted = get_untrusted_types(file=path)
if untrusted:
    raise RuntimeError(f"Unreviewed types: {untrusted}")
model = load(path, trusted=[])
prediction = model.predict(["Quero alugar uma casa em Franca"])[0]

Input and output

Input is a non-empty Portuguese string. Output is one label. predict_proba produces model_score, a relative model value that is not a calibrated probability or certainty. The local application applies a separately documented operational policy and routes uncertain cases to human review.

See example_input.json and example_output.json for synthetic examples.

Intended use

local educational demonstration;
reproducible comparison baseline;
study of traditional NLP for small synthetic datasets.

Out-of-scope use, limitations, and safety

Do not use this model for legal, financial, contractual, housing eligibility, tenant screening, emergency, or autonomous customer-service decisions. It was not trained on real conversations and may fail on dialects, spelling variation, ambiguous requests, code-switching, or unseen scenarios. Human escalation is required for uncertain or consequential cases.

No real messages or personal data are included. This model is not ready for production. Before any public or production use, review privacy, representative data, fairness, monitoring, security, dependency licenses, and incident handling.

Downloads last month: -