its-zion-18/music-text-distilbert-predictor

DistilBERT-based Music Era Classifier

This repository contains a fine-tuned text classification model based on distilbert-base-uncased. The model is designed to classify short text descriptions of eras in classical music into one of four historical musical eras: 0, 1, 2, and 3.

Model Architecture & Training

The model was trained using the Hugging Face Trainer API. It utilizes a distilbert-base-uncased pre-trained model with a classification head on top.

• —Tokenizer: AutoTokenizer.from_pretrained("distilbert-base-uncased")

• —Model: AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased")

• —Training Arguments: Learning Rate: 2×10−5

• —Epochs: 5

• —Batch Size: 8

• —Evaluation Strategy: Per epoch

• —Metric: accuracy

• —Optimizer: AdamW

music-text-distilbert-predictor

This model is a fine-tuned version of distilbert-base-uncased on the samder03/2025-24679-text-dataset. It achieves the following results on the evaluation set:

• —Loss: 0.0495
• —Accuracy: 1.0
• —F1: 1.0
• —Precision: 1.0
• —Recall: 1.0

Limitations

This model's primary limitations are:

Numerical Labels: The model outputs a numerical label (0, 1, 2, or 3). An external lookup table is required to map these numbers to their corresponding musical era names.

Language & Casing: As the model is based on distilbert-base-uncased, it is designed for English-language text and does not differentiate between uppercase and lowercase letters. It will not work for other languages.

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 2e-05
• —trainbatchsize: 8
• —evalbatchsize: 8
• —seed: 42
• —optimizer: Use OptimizerNames.ADAMWTORCHFUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• —lrschedulertype: linear
• —num_epochs: 5

Training results

Framework versions

• —Transformers 4.56.1
• —Pytorch 2.8.0+cu126
• —Datasets 4.0.0
• —Tokenizers 0.22.0

Potential Errors

There could be a problem with dataleakage because the accuracy is at 100% Because the model has already been trained on the augmented data, which is just a derivative of the original data, the original dataset isn't a true holdout set. The model is essentially being tested on data that it has already seen and, in some cases, memorized.