resume-section-classifier

Resume Section Classifier

A fine-tuned DistilBERT model that classifies resume text sections into 8 categories. Designed for automated resume parsing pipelines where incoming text needs to be segmented and labeled by section type.

Labels

Training Procedure

Data

The model was trained on a synthetic dataset generated programmatically using data_generator.py. The generator uses:

• Template-based generation with randomized entities (names, companies, universities, skills, dates, cities, etc.) drawn from curated pools of 30-50+ items each
• Structural variation across multiple formatting styles per category (e.g., education entries can appear as single-line, multi-line, with/without GPA, with activities, etc.)
• Data augmentation via synonym replacement (30% probability per eligible word) and formatting variation (whitespace, bullet styles, separators)
• Optional section headers prepended with 40% probability to teach the model to handle both headed and headless sections

Default configuration produces 1,920 examples (80 base examples x 3 variants x 8 categories), split 80/10/10 into train/validation/test sets with stratified sampling.

Hyperparameters

Training Infrastructure

• Fine-tuning takes approximately 5-10 minutes on a single GPU or 15-30 minutes on CPU
• Model size: ~67M parameters (DistilBERT base)
• Mixed precision (FP16) enabled automatically when CUDA is available

Metrics

These figures are INDICATIVE only. They are measured on a held-out split of the model's own synthetic, self-generated data (192 examples, stratified) — not a real-world benchmark. Because the train and test sets are produced by the same template-based generator, they share vocabulary and structure, so these scores overstate how the model will perform on real resumes.

Measured on the published checkpoint (4 epochs, seed 42, 192-example stratified synthetic test split):

A perfect score here is evidence of the leak described above, not of quality: the test split is generated by the same templates as the training split.

Out-of-distribution spot-check

To give the synthetic score some context, the checkpoint was run against eight hand-written snippets in the style of real resumes, one per section type, none drawn from the generator:

Both misses were low-confidence and collapsed toward the broader categories — an experience bullet read as summary (0.39), and a projects bullet read as skills (0.48). Treat ~0.75 on real text as the more realistic expectation, and treat predictions under ~0.5 confidence as unreliable.

Note: exact metrics depend on the random seed and training run.

Usage

Quick Start with Transformers Pipeline

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="gr8monk3ys/resume-section-classifier",
)

result = classifier(
    "Bachelor of Science in Computer Science, Stanford University, 2023. GPA: 3.9/4.0"
)
print(result)
# [{'label': 'education', 'score': 0.98}]

Full Resume Classification

from inference import ResumeSectionClassifier

classifier = ResumeSectionClassifier("gr8monk3ys/resume-section-classifier")

resume_text = """
John Smith
[email protected] | (415) 555-1234 | San Francisco, CA
linkedin.com/in/johnsmith | github.com/johnsmith

SUMMARY
Experienced software engineer with 5+ years building scalable web applications
and distributed systems. Passionate about clean code and mentoring junior developers.

EXPERIENCE
Senior Software Engineer | Google | San Francisco, CA | Jan 2021 - Present
- Led migration of monolithic application to microservices architecture
- Mentored 4 junior engineers through code reviews and pair programming
- Reduced API latency by 40% through caching and query optimization

Software Engineer | Stripe | San Francisco, CA | Jun 2018 - Dec 2020
- Built payment processing APIs handling 10M+ transactions daily
- Implemented CI/CD pipeline reducing deployment time from hours to minutes

EDUCATION
B.S. in Computer Science, Stanford University (2018)
GPA: 3.8/4.0 | Dean's List

SKILLS
Languages: Python, Java, Go, TypeScript
Frameworks: React, Django, Spring Boot, gRPC
Tools: Docker, Kubernetes, AWS, PostgreSQL, Redis
"""

analysis = classifier.classify_resume(resume_text)
print(analysis.summary())

Training from Scratch

# Install dependencies
pip install -r requirements.txt

# Generate data and train
python train.py --epochs 4 --batch-size 16

# Train and push to Hub
python train.py --push-to-hub --hub-model-id gr8monk3ys/resume-section-classifier

# Inference
python inference.py --file resume.txt
python inference.py --text "Python, Java, React, Docker" --single

Generating Custom Training Data

# Generate data with custom settings
python data_generator.py \
    --examples-per-category 150 \
    --augmented-copies 3 \
    --output data/resume_sections.csv \
    --print-stats

Project Structure

resume-section-classifier/
  data_generator.py    # Synthetic data generation with templates and augmentation
  train.py             # Full fine-tuning pipeline with HuggingFace Trainer
  inference.py         # Section splitting and classification API + CLI
  requirements.txt     # Python dependencies
  README.md            # This model card

Limitations

• Synthetic training data: The model is trained entirely on programmatically generated text. Performance on real resumes with unusual formatting, non-English content, or domain-specific jargon may be lower than synthetic test metrics suggest.
• English only: All training data is in English. The model will not reliably classify sections in other languages.
• Section granularity: The model classifies pre-split sections. The built-in text splitter uses heuristics (section headers, paragraph breaks) that may not correctly segment all resume formats (e.g., single-column vs. multi-column PDFs, tables).
• 8 categories only: Some resume sections (e.g., Publications, Volunteer Work, Languages, Interests, References) are not covered by the current label set. These will be assigned to the closest matching category.
• Max length: Input is truncated to 256 tokens. Very long sections may lose information.

Intended Use

• Automated resume parsing pipelines
• HR tech and applicant tracking systems
• Resume formatting and analysis tools
• Educational demonstrations of text classification fine-tuning

This model is not intended for making hiring decisions. It is a text classification tool for structural parsing only.

Author

Lorenzo Scaturchio (gr8monk3ys)