This model is fine-tuned from google-bert/bert-large-uncased for text classification tasks, specifically sentiment analysis of customer feedback on electrical devices - circuit breakers, transformers, smart meters, inverters, solar panels, power strips etc. The model has been optimized to classify sentiments into categories such as Positive, Negative, Neutral, and Mixed with high precision and recall, making it ideal for analyzing product reviews, customer surveys, and other feedback to derive actionable insights.

Training Data

The model was trained on the disham993/ElectricalDeviceFeedbackBalanced dataset, which has been carefully balanced to address class imbalances effectively. Original dataset which is imbalanced: disham993/ElectricalDeviceFeedback.

Model Details

Base Model: google-bert/bert-large-uncased
Task: text-classification
Language: en
Dataset: disham993/ElectricalDeviceFeedbackBalanced

Training procedure

Training hyperparameters

The model was fine-tuned using the following hyperparameters:

Evaluation Strategy: epoch
Learning Rate: 1e-5
Batch Size: 32 (for both training and evaluation)
Number of Epochs: 5
Weight Decay: 0.01

Evaluation results

The following metrics were achieved during evaluation:

F1 Score: 0.8928
Accuracy: 0.8897
eval_runtime: 2.5806
eval_samples_per_second: 523.905
eval_steps_per_second: 16.663

Usage

You can use this model for Sentiment Analysis of the Electrical Device Feedback as follows:

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification, pipeline

model_name = "disham993/electrical-classification-bert-large"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)
nlp = pipeline("text-classification", model=model, tokenizer=tokenizer)

text = "The new washing machine is efficient but produces a bit of noise."
classification_results = nlp(text)
print(classification_results)

Limitations and bias

The dataset includes synthetic data generated using Llama 3.1:8b, and despite careful optimization and prompt engineering, the model is not immune to errors in labeling. Additionally, as LLM technology is still in its early stages, there may be inherent inaccuracies or biases in the generated data that can impact the model's performance.

This model is intended for research and educational purposes only, and users are encouraged to validate results before applying them to critical applications.

Training Infrastructure

For a complete guide covering the entire process - from data tokenization to pushing the model to the Hugging Face Hub - please refer to the GitHub repository.

Last update

2025-01-05

Citation

@misc{modernbert,
      title={Smarter, Better, Faster, Longer: A Modern Bidirectional Encoder for Fast, Memory Efficient, and Long Context Finetuning and Inference}, 
      author={Benjamin Warner and Antoine Chaffin and Benjamin Clavié and Orion Weller and Oskar Hallström and Said Taghadouini and Alexis Gallagher and Raja Biswas and Faisal Ladhak and Tom Aarsen and Nathan Cooper and Griffin Adams and Jeremy Howard and Iacopo Poli},
      year={2024},
      eprint={2412.13663},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2412.13663}, 
}