Thraaxxxx/radarodio-v01-binary-ensemble

Radar Ódio v0.1 — Binary Production Ensemble

Frozen production artifact for binary hate-speech classification in Portuguese.

Architecture

• Backbone: pablocosta/bertabaporu-base-uncased
• Task: nao_hate / hate
• Final experimental condition: weighted
• Hyperparameter condition: hp_03
• Ensemble seeds: [13, 42, 97, 123, 2026]
• Ensemble aggregation: arithmetic mean of the five logits
• Calibration: temperature scaling
• Temperature: stored in config.json
• Final decision threshold: stored in config.json
• Truncation: Policy B, 255 first + 255 last content tokens
• Special tokens: ['<URL>', '<USER>', '<EMAIL>', '<PHONE>']

All production parameters are encapsulated in the model configuration.

Frozen v0.1 test result

• Macro-F1: 0.688161
• F1-hate: 0.478181

The model condition was selected before final test evaluation.

Production loading

from transformers import AutoTokenizer, AutoModelForSequenceClassification

repo = "Thraaxxxx/radarodio-v01-binary-ensemble"
revision = "<PIN_THE_COMMIT_SHA_PRINTED_BY_THE_UPLOAD_NOTEBOOK>"

tokenizer = AutoTokenizer.from_pretrained(
    repo,
    revision=revision,
    token=HF_TOKEN,
)

model = AutoModelForSequenceClassification.from_pretrained(
    repo,
    revision=revision,
    token=HF_TOKEN,
    trust_remote_code=True,
)

model.eval()

result = model.classify(
    "texto em português",
    tokenizer=tokenizer,
)

print(result)

The application does not need to know the temperature or binary threshold.

Output

{
    "label": "hate",
    "score": 0.91,
    "hate_probability": 0.91,
    "original_content_tokens": 37,
    "truncated_policy_b": False
}

Security

For production deployment, use a fine-grained Hugging Face token with read-only access and pin a specific repository commit.

Base model

BERTabaporu Base Uncased. Base model license: MIT.

Limitations

This model is a text-only classifier. Context outside the text, multimodal content, temporal drift and domain shift may affect predictions.