Nebulixlabs/Quantum-Classifier-Calibrated

Quantum-Clasifier-Finetuned

Fine-tuned version of:

Nebulixlabs/Quantum-Classifier

Architecture

• Parameters: 1,000,000
• Vocabulary: 4096
• Context length: 128
• Hidden size: 112
• Attention heads: 7
• Transformer layers: 4
• FFN: 336
• Classifier hidden: 170

Fine-tuning

Datasets:

• Nebulixlabs/safety-dataset
• KimDongH/spam_dataset-train-eval-test2

Batch size:

128

Target tokens:

30,000,000

Actual tokens:

30,010,756

Epochs:

9

Labels

0 = safe

1 = unsafe

Confidence

Temperature calibrated on validation data:

1.479935

Evaluation

The held-out test split, where available, is evaluated only after fine-tuning.

Multilingual smoke tests cover:

• English
• Hindi
• Hinglish