This repository contains all trained models for the Drug Use Modeler project: a binary text classifier that detects first-person drug use references in Spanish-language tweets. The models are trained and evaluated on the Drug Use Corpus (Spanish).

Overview

The experimental setup spans 48 configurations: 2 corpus variants × 3 preprocessing pipelines × 8 algorithms. The goal is to determine which combination best handles the ambiguity, irony, and obfuscation present in real-world social media drug discourse.

Corpus Variants

Variant	Description
`pre-filtered-corpus`	Cleaned dataset with noise and ambiguous tweets removed
`raw-corpus`	Full unfiltered dataset as collected from Twitter/X

Preprocessing Pipelines

Pipeline	Description
`standard`	Baseline tokenization and normalization
`irony`	Adds `[IRONIA]` token before sarcastic/ironic tweets
`obfuscated`	Named entities (substances, slang) masked via NER

Models

Model	Type	Representation
`naive_bayes`	Multinomial Naive Bayes	TF-IDF (5,000 features, bigrams)
`logistic_regression`	Logistic Regression	TF-IDF (5,000 features, bigrams)
`svm`	Linear SVC	TF-IDF (5,000 features, bigrams)
`random_forest`	Random Forest	TF-IDF (5,000 features, bigrams)
`ffn`	Feed-Forward Network (100→64→32→1)	Word2Vec 100-dim
`cnn`	TextCNN (filters 3/4/5)	Word2Vec 100-dim
`rnn`	BiLSTM (hidden=64, bidirectional)	Word2Vec 100-dim
`bert_base`	Fine-tuned BETO	`dccuchile/bert-base-spanish-wwm-cased`

Repository Structure

├── pre-filtered-corpus/
│   ├── naive_bayes/{standard,irony,obfuscated}/
│   │   ├── model.joblib
│   │   └── vectorizer.joblib
│   ├── logistic_regression/{standard,irony,obfuscated}/   (same structure)
│   ├── svm/{standard,irony,obfuscated}/                   (same structure)
│   ├── random_forest/{standard,irony,obfuscated}/         (same structure)
│   ├── ffn/{standard,irony,obfuscated}/
│   │   └── model.pt
│   ├── cnn/{standard,irony,obfuscated}/
│   │   └── model.pt
│   ├── rnn/{standard,irony,obfuscated}/
│   │   └── model.pt
│   ├── word2vec/{standard,irony,obfuscated}/
│   │   └── word2vec.model
│   └── bert_base/{standard,irony,obfuscated}/
│       ├── model/        (HuggingFace saved model)
│       └── tokenizer/    (HuggingFace tokenizer)
└── raw-corpus/           (same structure)

Note: For raw-corpus, Naive Bayes, SVM, and Random Forest models were not persisted — re-train inline with identical hyperparameters (seed=42) using the provided training splits.

Results

All metrics are macro-averaged over a balanced test set (225 NEGATIVE / 225 POSITIVE, 450 total), stratified jointly by class and substance. BETO's checkpoint is selected on the validation split (not test), and test is scored exactly once per run — see Iteration 3 in the source repo's evaluation log for the full methodology and history across iterations.

Pre-filtered Corpus

Standard

Model	Accuracy	Precision	Recall	F1
BERT (Base)	86.00%	86.01%	86.00%	86.00%
Logistic Regression	80.22%	80.29%	80.22%	80.21%
SVM	79.78%	79.81%	79.78%	79.77%
Naive Bayes	77.33%	77.39%	77.33%	77.32%
CNN	77.33%	77.47%	77.33%	77.30%
Random Forest	76.67%	76.67%	76.67%	76.67%
RNN (BiLSTM)	76.67%	76.71%	76.67%	76.66%
FFN	72.89%	72.98%	72.89%	72.86%

Irony

Model	Accuracy	Precision	Recall	F1
BERT (Base)	85.33%	85.34%	85.33%	85.33%
Logistic Regression	80.44%	80.50%	80.44%	80.43%
CNN	79.33%	79.40%	79.33%	79.32%
SVM	79.11%	79.12%	79.11%	79.11%
Random Forest	79.11%	79.12%	79.11%	79.11%
RNN (BiLSTM)	78.44%	78.51%	78.44%	78.43%
Naive Bayes	77.78%	77.83%	77.78%	77.77%
FFN	76.67%	76.68%	76.67%	76.66%

Obfuscated

Model	Accuracy	Precision	Recall	F1
BERT (Base)	85.78%	85.80%	85.78%	85.78%
SVM	79.78%	79.81%	79.78%	79.77%
Logistic Regression	78.89%	79.02%	78.89%	78.87%
CNN	78.22%	78.28%	78.22%	78.21%
Naive Bayes	76.89%	76.92%	76.89%	76.88%
Random Forest	76.44%	76.52%	76.44%	76.43%
FFN	75.78%	75.78%	75.78%	75.78%
RNN (BiLSTM)	75.11%	75.14%	75.11%	75.10%

Raw Corpus

Standard

Model	Accuracy	Precision	Recall	F1
BERT (Base)	85.56%	85.71%	85.56%	85.54%
Logistic Regression	80.67%	80.74%	80.67%	80.66%
SVM	80.00%	80.01%	80.00%	80.00%
Naive Bayes	79.78%	79.83%	79.78%	79.77%
CNN	79.78%	79.99%	79.78%	79.74%
RNN (BiLSTM)	79.11%	79.51%	79.11%	79.04%
FFN	78.00%	78.16%	78.00%	77.97%
Random Forest	74.44%	74.50%	74.44%	74.43%

Irony

Model	Accuracy	Precision	Recall	F1
BERT (Base)	87.56%	87.60%	87.56%	87.55%
Logistic Regression	80.67%	80.77%	80.67%	80.65%
SVM	80.00%	80.01%	80.00%	80.00%
Naive Bayes	79.78%	79.83%	79.78%	79.77%
CNN	79.33%	79.50%	79.33%	79.30%
RNN (BiLSTM)	77.56%	77.96%	77.56%	77.47%
Random Forest	76.00%	76.02%	76.00%	76.00%
FFN	74.67%	76.30%	74.67%	74.27%

Obfuscated

Model	Accuracy	Precision	Recall	F1
BERT (Base)	87.78%	87.80%	87.78%	87.78%
Logistic Regression	81.78%	81.84%	81.78%	81.77%
SVM	80.22%	80.24%	80.22%	80.22%
Naive Bayes	80.22%	80.25%	80.22%	80.22%
RNN (BiLSTM)	80.22%	80.29%	80.22%	80.21%
CNN	80.00%	80.01%	80.00%	80.00%
FFN	77.78%	77.80%	77.78%	77.77%
Random Forest	76.22%	76.37%	76.22%	76.19%

Best Models Summary

Corpus	Best Configuration	F1
Pre-filtered	BERT (Base) — Standard	86.00%
Raw	BERT (Base) — Obfuscated	87.78% (best overall)

Note: Naive Bayes, SVM, and Random Forest are not the top model in any raw-corpus or pre-filtered-corpus variant as of this iteration — BERT (Base) leads across all six corpus/variant combinations. See the evaluation log for the full iteration history and how this ranking changed as split/evaluation bugs were fixed.

How to Use

Sklearn models (Naive Bayes, Logistic Regression, SVM, Random Forest)

python

import joblib
from huggingface_hub import hf_hub_download

# Download model and vectorizer
model_path = hf_hub_download(
    repo_id="lhbelfanti/drug-use-models",
    filename="pre-filtered-corpus/svm/standard/model.joblib"
)
vectorizer_path = hf_hub_download(
    repo_id="lhbelfanti/drug-use-models",
    filename="pre-filtered-corpus/svm/standard/vectorizer.joblib"
)

model = joblib.load(model_path)
vectorizer = joblib.load(vectorizer_path)

texts = ["me tomé una línea antes de salir"]
X = vectorizer.transform(texts)
print(model.predict(X))  # [1] → POSITIVE

BERT (BETO)

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
from huggingface_hub import snapshot_download
import torch

# Download model and tokenizer
model_dir = snapshot_download(
    repo_id="lhbelfanti/drug-use-models",
    allow_patterns="pre-filtered-corpus/bert_base/standard/**"
)

tokenizer = AutoTokenizer.from_pretrained(f"{model_dir}/pre-filtered-corpus/bert_base/standard/tokenizer")
model = AutoModelForSequenceClassification.from_pretrained(f"{model_dir}/pre-filtered-corpus/bert_base/standard/model")
model.eval()

text = "me tomé una línea antes de salir"
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=128)
with torch.no_grad():
    logits = model(**inputs).logits
pred = torch.argmax(logits, dim=1).item()
print("POSITIVE" if pred == 1 else "NEGATIVE")

Training Details

Dataset: lhbelfanti/drug-use-corpus
Split: 70% train / 15% val / 15% test, stratified jointly by class (POSITIVE/NEGATIVE) and substance (Cocaína/Marihuana/Heroína/Ecstasy), seed=42. Cross-split text duplicates are resolved after splitting.
TF-IDF: max_features=5000, ngram_range=(1,2)
Word2Vec: 100 dimensions, window=5, min_count=1, trained in-domain
BERT base: dccuchile/bert-base-spanish-wwm-cased, fine-tuned 3 epochs, lr=2e-5, batch=16. Checkpoint selected on the validation split; test is evaluated once, after training.
Framework: scikit-learn 1.5, PyTorch 2.2.2, Transformers 4.40.2

Related Resources

Dataset: lhbelfanti/drug-use-corpus
Code: github.com/lhbelfanti/drug-use-modeler

License

MIT