language: en
license: mit
library_name: transformers
tags:
- text-classification
- pytorch
- distilbert
- customer-support
- nlp
datasets:
- Bitext/customer-support-intent-dataset
metrics:
- accuracy
- f1
pipeline_tag: text-classification
DistilBERT for Automated Tech-Support Classification
This model is a fine-tuned version of DistilBERT (distilbert-base-uncased) trained to classify customer support tickets into 27 specific intents across 11 major categories.
This model is the "Brain" of the Automated Tech-Support Ticketing System project.
🚀 Model Details
- Architecture: DistilBERT (Transformers)
- Task: Multi-class Text Classification
- Intents: 27 (e.g.,
cancel_order,recover_password,edit_account, etc.) - Framework: PyTorch & Hugging Face Transformers
📊 Performance (Week 2 Results)
The model achieved near-perfect scores on the Bitext Customer Support Dataset:
- Training Accuracy: 100.00%
- Validation Accuracy: 99.76%
- Macro Average F1-Score: 1.00
📂 Artifacts in this Repo
best_model_state.bin: The trained PyTorch model weights.tokenizer/: Full configuration for the BERT tokenizer.label_encoder.joblib: The mapping for the 27 intent classes.
🛠️ Integration with Project
This model is designed to be used in conjunction with a FastAPI backend and a Gemini 2.5-flash reasoning layer.
To use this model in your local setup, you can clone this repository or use the huggingface_hub library to download the artifacts into the models/ directory of the main project.
How to Load (Example):
from transformers import DistilBertForSequenceClassification, DistilBertTokenizer
import torch
# Path to the downloaded model
model = DistilBertForSequenceClassification.from_pretrained("./models/tokenizer", num_labels=27)
model.load_state_dict(torch.load("./models/best_model_state.bin"))
🔗 Main Project Repository
For the full end-to-end implementation (FastAPI, Streamlit, and LLM Integration), please visit my GitHub: 👉 GitHub Repository
Developed by Baltasar Patrizhard Djata Part of the "Automated Tech-Support Ticketing System" Portfolio Project (2026).