Dc-4nderson / tone-classifier

DistilBERT Tone Classification Model

This model fine-tunes distilbert-base-uncased to classify tone into 8 categories relevant to community and mentorship transcripts.

📌 Labels

uplifting

thoughtful

practical

reflective

motivational

informative

neutral

negative

📊 Dataset

The model is trained on the tone-dataset , a dataset containing 1000+ labeled examples created for the MyVillageProject tone classification task. Data includes first-person and third-person statements, anecdotes, factual notes, and reflective entries.

🚀 Training

Base model: distilbert-base-uncased

Optimizer: AdamW (lr=5e-5)

Batch size: 16

Epochs: 8

Loss: CrossEntropy

Metrics: Accuracy + Weighted F1

📈 Validation Metrics Epoch Training Loss Validation Loss Accuracy F1 1 No log 0.484719 0.894161 0.895220 2 No log 0.264668 0.923358 0.923200 3 No log 0.243101 0.930657 0.930599 4 No log 0.302434 0.916058 0.918166 5 No log 0.305320 0.923358 0.923836 6 No log 0.294621 0.916058 0.916176 7 No log 0.303021 0.919708 0.919583 8 0.215900 0.298230 0.916058 0.915722

Final Training Summary:

TrainOutput(global_step=552, training_loss=0.1959800598198089, metrics={ 'train_runtime': 39.2397, 'train_samples_per_second': 223.244, 'train_steps_per_second': 14.067, 'total_flos': 290134644572160.0, 'train_loss': 0.1959800598198089, 'epoch': 8.0 })

💻 Usage from transformers import pipeline

classifier = pipeline("text-classification", model="Dc-4nderson/tone-distilbert")

text = "Ronnie mentioned the turnout was twice what they expected, and it felt like a victory." print(classifier(text))

Output:

[{'label': 'uplifting'}]

🔖 License

Apache-2.0

👥 Maintainer

Dequan Anderson/ Dc-4nderson