0xnbk/resume-domain-triplets-train-v1-en

Resume-Domain Triplets Dataset v1 (English)

Dataset Description

resume-domain-triplets-train-v1-en is a contrastive learning dataset designed for training domain classification models using triplet loss. This dataset enables fine-tuning of sentence transformers to understand professional domain alignment between job descriptions, making it essential for building intelligent ATS (Applicant Tracking System) applications.

Key Features

• 📊 40K triplets (32K train, 8K validation)
• 🎯 13 professional domains with balanced representation
• 🔄 Triplet format ready for contrastive learning (anchor, positive, negative)
• ✅ Hard negative mining for effective domain separation
• 🌍 Multi-domain coverage across Technology, Healthcare, Finance, Education, and more
• 🔥 High quality triplets mined from real LinkedIn job postings

Dataset Structure

Data Format

Each triplet contains:

Column	Type	Description
anchor	string	Job description used as anchor point
positive	string	Job description from the same domain as anchor
negative	string	Job description from a different domain (hard negative)
anchor_domain	string	Professional domain of anchor
positive_domain	string	Professional domain of positive (same as anchor)
negative_domain	string	Professional domain of negative (different from anchor)

Data Splits

Split	Triplets	Percentage
Train	32,000	80%
Validation	8,000	20%
Total	40,000	100%

Example Triplets

Technology Domain:

Anchor: "Senior Software Engineer with Python, Django, and AWS experience..."
Positive: "Backend Developer position requiring Node.js and cloud deployment..."
Negative: "Registered Nurse position in ICU requiring patient care skills..."

Anchor Domain: technology
Positive Domain: technology
Negative Domain: healthcare

Healthcare Domain:

Anchor: "Registered Nurse with 5+ years ICU experience..."
Positive: "Clinical Nurse Specialist for cardiac care unit..."
Negative: "Financial Analyst requiring CFA and portfolio management..."

Anchor Domain: healthcare
Positive Domain: healthcare
Negative Domain: finance

Professional Domains

The dataset covers 13 professional domains with balanced representation:

Domain	Examples	Approximate Count
Technology	Software Engineering, Data Science, DevOps, AI/ML	~2,600
Healthcare	Nursing, Medical Practice, Clinical Research	~2,400
Finance	Banking, Investment, Financial Analysis, Fintech	~2,600
Education	Teaching, Academic Research, EdTech	~2,600
Sales & Marketing	Digital Marketing, Business Development, SEO	~2,600
Human Resources	Recruitment, HR Management, Training	~2,600
Legal	Legal Practice, Compliance, Contracts	~2,600
Manufacturing & Operations	Production, Quality Control, Supply Chain	~2,600
Retail & Hospitality	Customer Service, Store Management	~2,600
Construction & Real Estate	Project Management, Property Development	~2,600
Government & Nonprofit	Public Service, Community Programs	~2,600
Media & Entertainment	Content Creation, Broadcasting, Production	~2,500
Design	Graphic Design, UX/UI, Creative Direction	~2,400

Source Data

This dataset was generated from the LinkedIn Jobs Dataset (arjunth2001/linkedin-jobs-postings-and-skills-2024), which contains real job postings from LinkedIn.

Data Generation Process

1. Domain Extraction: Job postings classified into 13 professional domains
2. Triplet Mining:
   • Anchor: Random job description from a domain
   • Positive: Different job from the same domain
   • Negative: Job from a different domain (hard negative sampling)
3. Balancing: Equal representation across all domains
4. Quality Control: Filtered for text quality, length, and domain diversity
5. Validation Split: 20% held out for model evaluation

Hard Negative Mining Strategy

Negative examples are carefully selected to be challenging:

• High semantic similarity but different domains (e.g., "Data Scientist" vs "Data Analyst in Finance")
• Cross-domain skills overlap (e.g., "Project Manager in Tech" vs "Project Manager in Construction")
• Similar job levels but different industries

Intended Use

Primary Use Cases

1. Domain Classification Model Training: Fine-tune sentence transformers for domain detection
2. Contrastive Learning: Train embeddings that cluster same-domain jobs together
3. ATS Domain Matching: Improve resume-job matching by understanding domain alignment
4. Cross-Domain Transfer: Study patterns in cross-domain skill transferability

Model Training

This dataset is designed for training with TripletLoss using sentence transformers:

Recommended Base Models:

• jinaai/jina-embeddings-v2-base-en
• sentence-transformers/all-MiniLM-L6-v2
• sentence-transformers/all-mpnet-base-v2
• Any sentence transformer model

Expected Performance: Models trained on this dataset typically achieve >85% domain classification accuracy with strong cross-domain separation.

Example Training Code

from sentence_transformers import SentenceTransformer, losses, InputExample
from torch.utils.data import DataLoader
from datasets import load_dataset
import pandas as pd

# Load dataset
dataset = load_dataset("0xnbk/resume-domain-triplets-train-v1-en")
train_df = pd.DataFrame(dataset['train'])

# Prepare triplets
train_examples = []
for _, row in train_df.iterrows():
    train_examples.append(
        InputExample(texts=[row['anchor'], row['positive'], row['negative']])
    )

# Load pre-trained base model (e.g., nbk-ats-semantic-v1-en)
model = SentenceTransformer('0xnbk/nbk-ats-semantic-v1-en')

# Define loss and dataloader
train_dataloader = DataLoader(train_examples, shuffle=True, batch_size=16)
train_loss = losses.TripletLoss(
    model=model,
    distance_metric=losses.TripletDistanceMetric.COSINE,
    triplet_margin=0.5
)

# Train
model.fit(
    train_objectives=[(train_dataloader, train_loss)],
    epochs=3,
    warmup_steps=100,
    optimizer_params={'lr': 5e-6},
    output_path='./domain-model'
)

# Save
model.save('./domain-model')

Inference Example

from sentence_transformers import SentenceTransformer
from scipy.spatial.distance import cosine

# Load trained model
model = SentenceTransformer('./domain-model')

# Test domain similarity
resume = "Software engineer with 5 years Python and Django experience"
job_tech = "Senior Python developer position"
job_healthcare = "Registered nurse position in ICU"

# Encode
resume_emb = model.encode(resume)
tech_emb = model.encode(job_tech)
healthcare_emb = model.encode(job_healthcare)

# Calculate similarities
tech_similarity = 1 - cosine(resume_emb, tech_emb)
healthcare_similarity = 1 - cosine(resume_emb, healthcare_emb)

print(f"Technology match: {tech_similarity:.3f}")  # High score (same domain)
print(f"Healthcare match: {healthcare_similarity:.3f}")  # Low score (different domain)

Dataset Statistics

Size Metrics

• Total size: ~280MB (CSV format with triplet text fields)
• Average text length: ~800 characters per job description
• Token count: ~35M tokens (estimated with BERT tokenizer)
• Triplets per domain: ~3,000 average (balanced distribution)

Domain Balance

All 13 domains have approximately equal representation (2,400-2,600 triplets per domain), ensuring no domain bias in trained models.

Training Details

Model: nbk-ats-domain-v1-en

This dataset was used to train the nbk-ats-domain-v1-en model with the following configuration:

• Base Model: nbk-ats-semantic-v1-en (Jina embeddings fine-tuned for ATS semantic matching)
• Loss Function: TripletLoss with cosine distance
• Triplet Margin: 0.5
• Epochs: 3
• Batch Size: 8 (with gradient accumulation: effective batch size 128)
• Learning Rate: 5e-6
• Hardware: NVIDIA A6000 48GB GPU
• Training Time: ~2 hours

Performance Metrics

• Domain Classification Accuracy: >85%
• Cross-Domain Separation: Strong embedding clusters
• Same-Domain Similarity: High cosine similarity (>0.7)
• Cross-Domain Similarity: Low cosine similarity (<0.3)

Limitations and Considerations

Known Limitations

1. Domain Boundaries: Some domains have natural overlap (e.g., "Technology" and "Data Science")
2. Job Description Focus: Dataset contains job descriptions, not resumes (domain patterns differ)
3. Language: Currently only English language support
4. Domain Granularity: 13 broad domains may not capture niche specializations
5. Temporal Bias: Reflects job market terminology as of 2024-2025

Ethical Considerations

• Bias: Domain classification may reflect societal biases present in job postings
• Privacy: No personally identifiable information (PII) included
• Fairness: Users should validate model fairness across protected characteristics
• Transparency: Triplets derived from public LinkedIn job postings

Citation

If you use this dataset in your research or applications, please cite:

@dataset{resume_domain_triplets_v1,
  author = {NBK},
  title = {Resume-Domain Triplets Dataset v1 (English)},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/0xnbk/resume-domain-triplets-train-v1-en}
}

Source Dataset Citation

This dataset is derived from the LinkedIn Jobs Dataset:

@dataset{linkedin_jobs_2024,
  author = {Arjun Thangaraju},
  title = {LinkedIn Jobs Postings and Skills 2024},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/arjunth2001/linkedin-jobs-postings-and-skills-2024}
}

Model Citation

If you use the model trained on this dataset:

@model{nbk_ats_domain_v1,
  author = {NBK},
  title = {NBK ATS Domain Classifier v1 (English)},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/0xnbk/nbk-ats-domain-v1-en}
}

License

This dataset is released under the Apache 2.0 License.

Copyright 2025 NBK (nbk.dev)

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

Updates and Maintenance

• Version: 1.0.0
• Last Updated: October 2025
• Maintained by: NBK (nbk.dev)
• Issues: Report issues on the dataset discussion page

Related Resources

• Source Dataset: arjunth2001/linkedin-jobs-postings-and-skills-2024
• Trained Model: 0xnbk/nbk-ats-domain-v1-en
• Base Model: 0xnbk/nbk-ats-semantic-v1-en
• Binary Classifier Dataset: 0xnbk/resume-domain-classifier-v1-en
• Application: LOCAL ATS - Privacy-first ATS Resume Analyzer

Contact

For questions, suggestions, or collaboration opportunities:

• GitHub: 0xnbk/localATS
• HuggingFace: @0xnbk
• Website: nbk.dev