README.md

---
language:
- en
base_model:
- distilbert/distilbert-base-uncased
license: apache-2.0
metrics:
- accuracy
- precision
- recall
- f1
library_name: adapter-transformers
tags:
- resume-classification
- multi-label-classification
- human-resources
- transformers
- distilbert
- career-guidance
- fine-tuned
---
# **Res-BERT**
Fine-tuned DistilBERT model for multi-label resume classification.

---

## **Model Overview**
**Res-BERT** is a fine-tuned version of the DistilBERT base model, trained on a multi-labeled dataset of resumes (`resume_corpus`) with occupation labels. This model can classify resumes into multiple occupation categories, making it a useful tool for HR teams, recruitment platforms, and AI-powered career assistants.

### **Base Model**
- **DistilBERT (uncased):** A smaller, faster, and cheaper version of BERT, pretrained on BookCorpus and English Wikipedia. It provides a balance of performance and efficiency for NLP tasks.

---

## **Dataset**
The **resume_corpus** dataset was used for training. It consists of resumes labeled with occupations. The dataset includes:
- **`resumes_corpus.zip`:** A collection of `.txt` files (resumes) with corresponding labels in `.lab` files.
- **`resumes_sample.zip`:** A consolidated text file, where each line contains:
  - Resume ID
  - Occupation labels (separated by `;`)
  - Resume text.
- **`normalized_classes`:** Associations between raw and normalized occupation labels.

### **Dataset Citation**
Jiechieu, K.F.F., Tsopze, N. *Skills prediction based on multi-label resume classification using CNN with model predictions explanation.* Neural Comput & Applic (2020). [DOI:10.1007/s00521-020-05302-x](https://doi.org/10.1007/s00521-020-05302-x).

---


## **Training Procedure**
The model was fine-tuned using:
- **Input Format:** Lowercased, tokenized text using WordPiece with a vocabulary of 30,000 tokens.
- **Hyperparameters:** Default settings of the Hugging Face `Trainer` API for DistilBERT-based sequence classification.
- **Preprocessing:**
  - Masking: 15% of tokens were masked during pretraining.
  - Split: 80% training, 10% validation, 10% test.
- **Hardware:** 8 16GB V100 GPUs, trained for 10 hours.

---

## **Intended Use**
### **Applications**
- Resume screening for recruitment platforms.
- Career guidance and job-matching services.
- Analyzing skills and experiences from resumes.

### **How to Use**
#### Using Transformers' pipeline:
```python
from transformers import pipeline

classifier = pipeline("text-classification", model="Res-BERT", tokenizer="Res-BERT", multi_label=True)
resumes = ["Software developer with 5 years of experience in Java and Python."]
predictions = classifier(resumes)

print(predictions)
```

#### Using Transformers' pipeline:
```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification

tokenizer = AutoTokenizer.from_pretrained("Res-BERT")
model = AutoModelForSequenceClassification.from_pretrained("Res-BERT")

text = "Experienced mechanical engineer with expertise in CAD and manufacturing."
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)

print(outputs.logits)

```
## Citations
@article{Sanh2019DistilBERTAD,
  title={DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter},
  author={Victor Sanh and Lysandre Debut and Julien Chaumond and Thomas Wolf},
  journal={ArXiv},
  year={2019},
  volume={abs/1910.01108}
}

@article{Jiechieu2020ResumeClassification,
  title={Skills prediction based on multi-label resume classification using CNN with model predictions explanation},
  author={K.F.F. Jiechieu and N. Tsopze},
  journal={Neural Comput & Applic},
  year={2020},
  doi={10.1007/s00521-020-05302-x}
}